WO2021029264A1

WO2021029264A1 - End point detecting device and end point detecting method

Info

Publication number: WO2021029264A1
Application number: PCT/JP2020/029746
Authority: WO
Inventors: 松尾　尚典
Original assignee: 株式会社荏原製作所
Priority date: 2019-08-09
Filing date: 2020-08-04
Publication date: 2021-02-18
Also published as: JP2021028099A

Abstract

This end point detecting device has a trained model obtained by machine learning of waveforms of measurement data between the start of polishing and the end of polishing that have been output during polishing in the past from each of a plurality of types of end point detecting sensors provided in one polishing unit. The end point detecting device is provided with a determination unit that receives, as an input, measurement data between the start of polishing and the current point of time that have been newly output during polishing from each of the plurality of types of end point detecting sensors, and estimates, as an output, whether the current point of time is the timing for an end point indicating the end of polishing.

Description

End point detection device, end point detection method

The present disclosure relates to an end point detection device and an end point detection method.

In recent years, as the integration of semiconductor devices has progressed, the wiring of circuits has become finer and the distance between wirings has become narrower. In the manufacture of semiconductor devices, many types of materials are repeatedly formed in a film shape on a silicon wafer to form a laminated structure. In order to form this laminated structure, a technique for flattening the surface of the wafer is important. As a means for flattening the surface of such a wafer, a polishing apparatus (also referred to as a chemical mechanical polishing apparatus) that performs chemical mechanical polishing (CMP) is widely used.

This type of polishing device generally holds a polishing table equipped with a polishing pad for polishing an object to be polished (a substrate such as a wafer) and a wafer to hold the object to be polished and press it against the polishing pad. Has a top ring to polish. The polishing table and the top ring are each rotationally driven by a drive unit (for example, a motor). Further, the polishing apparatus includes a nozzle for supplying a polishing liquid onto the polishing pad. While supplying the polishing liquid from the nozzle onto the polishing pad, the wafer is pressed against the polishing pad by the top ring, and the top ring and the polishing table are moved relative to each other to polish the wafer and flatten its surface.

In the polishing device, if the object to be polished is not sufficiently polished, the insulation between the circuits cannot be obtained and there is a risk of short circuit. Further, in the case of overpolishing, there arises a problem that the resistance value increases due to the decrease in the cross-sectional area of the wiring, or the wiring itself is completely removed and the circuit itself is not formed. Therefore, the polishing apparatus is required to detect the optimum polishing end point.

As one of the polishing end point detecting means, a method of detecting a change in polishing friction force when polishing is transferred to a substance of a different material is known. The semiconductor wafer, which is the object to be polished, has a laminated structure made of different materials such as a semiconductor, a conductor, and an insulator, and has a different coefficient of friction between layers of different materials. Therefore, this is a method of detecting a change in polishing friction force caused by the transfer of polishing to a different material layer. According to this method, the end point of polishing is when the polishing reaches a different material layer.

Further, the polishing apparatus can also detect the polishing end point by detecting the change in the polishing friction force when the polishing surface of the object to be polished becomes flat from the non-flat state.

Here, the polishing frictional force generated when polishing the object to be polished appears as a drive load of a drive unit that rotationally drives the polishing table or top ring. For example, when the drive unit is an electric motor, the drive load (torque) can be measured as a current flowing through the motor. Therefore, the motor current (torque current) can be detected by the current sensor, and the end point of polishing can be detected based on the change in the detected motor current.

As another example of the polishing end point detecting means, a method of detecting a change in a physical quantity of a semiconductor wafer by using an optical sensor or an eddy current type sensor incorporated in a polishing table is known.

Torque fluctuation detection (motor current fluctuation measurement) is excellent in detecting the end point of the part where the film quality of the sample to be polished changes. The optical method is excellent in detecting the amount of residual film of an insulating film such as an interlayer insulating film (ILD) and STI (Shallow Trench Isolation) and detecting the end point by the detection. The eddy current method is excellent in detecting the end point at the time when, for example, the plated metal film is polished to the lower insulating film which is the end point.

Japanese Unexamined Patent Publication No. 2018-58197 proposes to use a plurality of types of end point detection sensors in combination.

However, when multiple types of end point detection sensors are used in combination, which end point detection sensor measurement data should be prioritized and when the end point detection sensors should be prioritized by individually tuning each polishing unit. It was necessary for the worker to explicitly instruct whether to switch the order, etc., and there was a lot of work for individual response, which required time and cost for individual response. In addition, there is insufficient accuracy for detecting the end point of a high-precision requirement corresponding to a fine pattern.

It is desired to provide an end point detection device and an end point detection method that can improve the accuracy of end point detection.

The end point detection device according to one aspect of the present disclosure is
A trained model (for example, a tuned neural network system) that machine-learns the waveform of the measurement data from the start of polishing to the end of polishing output from each of the multiple types of end point detection sensors provided in one polishing unit during past polishing. ), And by inputting the measurement data from the start of polishing to the present time output from each of the plurality of types of end point detection sensors at the time of new polishing, it is estimated whether or not the present time is the timing of the end point indicating the end of polishing. It is provided with a determination unit that outputs the data.

The end point detecting device according to another aspect of the present disclosure is
A trained model (for example, a tuned neural network system) that machine-learns the waveform of the measurement data from the start of polishing to the end of polishing output from each of the multiple types of end point detection sensors provided in one polishing unit during past polishing. ), And the remaining time from the present time to the end point timing indicating the end of polishing is estimated by inputting the measurement data from the start of polishing to the present time output from each of the plurality of types of end point detection sensors at the time of new polishing. It is equipped with a determination unit that outputs data.

FIG. 1 is a plan view showing the overall configuration of the substrate processing apparatus according to the embodiment. FIG. 2 is a perspective view schematically showing the first polishing unit. FIG. 3 is a cross-sectional view schematically showing an example of the structure of the top ring. FIG. 4 is a cross-sectional view schematically showing another example of the structure of the top ring. FIG. 5 is a cross-sectional view for explaining a mechanism for rotating and swinging the top ring. FIG. 6 is a cross-sectional view schematically showing the internal structure of the polishing table. FIG. 7 is a schematic view for explaining an optical sensor provided on the polishing table. FIG. 8 is a schematic view for explaining a microwave sensor provided on the polishing table. FIG. 9 is a block diagram showing a configuration of an end point detection unit (end point detection device) according to an embodiment. FIG. 10A is a schematic diagram for explaining an example of the configuration of the trained model in one form. FIG. 10B is a schematic diagram for explaining an example of the configuration of the trained model in another form. FIG. 11A is a schematic diagram for explaining a modified example of the configuration of the trained model in one form. FIG. 11B is a schematic diagram for explaining a modified example of the configuration of the trained model in another form. FIG. 12 is an image diagram for explaining the learning content of the trained model. FIG. 13 is a diagram for explaining an example of processing of the determination unit for real-time measurement data during polishing. FIG. 14 is a diagram for explaining an example of processing of the determination unit for real-time measurement data during polishing. FIG. 15 is a diagram for explaining an example of processing of the determination unit for real-time measurement data during polishing. FIG. 16 is a diagram for explaining an example of processing of the timing adjustment unit. FIG. 17 is a diagram for explaining an example of processing of the timing adjustment unit. FIG. 18A is a flowchart showing an example of the end point detection method according to the embodiment. FIG. 18B is a flowchart showing another example of the end point detection method according to the embodiment. FIG. 19 is a diagram showing overall control by the control unit. FIG. 20 is a diagram showing a configuration according to an embodiment. FIG. 21 is a diagram showing a modified example of the configuration according to one embodiment. FIG. 22 is a diagram showing a modified example of the configuration according to one embodiment. FIG. 23 is a diagram showing a modified example of the configuration according to the embodiment.

The end point detection device according to the first aspect of the embodiment is
A trained model (for example, a tuned neural network system) that machine-learns the waveform of the measurement data from the start of polishing to the end of polishing output from each of the multiple types of end point detection sensors provided in one polishing unit during past polishing. ), And by inputting the measurement data from the start of polishing to the present time output from each of the plurality of types of end point detection sensors at the time of new polishing, it is estimated whether or not the present time is the timing of the end point indicating the end of polishing. It is provided with a determination unit for outputting.

According to such an aspect, which end point detection sensor's measurement data is preferentially used for the measurement data output from each of the plurality of types of end point detection sensors provided in one polishing unit at the time of new polishing. , The waveform of the measurement data from the start of polishing to the end of polishing output during past polishing was machine-learned without the operator explicitly instructing when to switch the priority between the end point detection sensors. By using a trained model (for example, a tuned neural network system), it is estimated and output whether or not the current time is the timing of the end point in view of the similarity with the waveform of the measurement data at the time of past polishing. can do. Therefore, it is possible to optimally combine and use the measurement data of a plurality of types of end point detection sensors, and it is possible to improve the accuracy of end point detection.

The end point detecting device according to the second aspect of the embodiment is
A trained model (for example, a tuned neural network system) that machine-learns the waveform of the measurement data from the start of polishing to the end of polishing output from each of the multiple types of end point detection sensors provided in one polishing unit during past polishing. ), And the remaining time from the present time to the end point timing indicating the end of polishing is estimated by inputting the measurement data from the start of polishing to the present time output from each of the plurality of types of end point detection sensors at the time of new polishing. It is equipped with a determination unit that outputs data.

According to such an aspect, which end point detection sensor's measurement data is preferentially used for the measurement data output from each of the plurality of types of end point detection sensors provided in one polishing unit at the time of new polishing. , The waveform of the measurement data from the start of polishing to the end of polishing output during past polishing was machine-learned without the operator explicitly instructing when to switch the priority between the end point detection sensors. By using a trained model (for example, a tuned neural network system), the remaining time from the present time to the timing of the end point indicating the end of polishing is estimated in consideration of the similarity with the waveform of the measurement data at the time of past polishing. Can be output. Therefore, it is possible to optimally combine and use the measurement data of a plurality of types of end point detection sensors, and it is possible to improve the accuracy of end point detection.

The end point detecting device according to the third aspect of the embodiment is an end point detecting device according to the first aspect.
A first polishing stop unit is further provided, which transmits a control signal for stopping polishing to the polishing unit when it is estimated by the determination unit that the current time is the timing of the end point.

The end point detecting device according to the fourth aspect of the embodiment is an end point detecting device according to the second aspect.
A first polishing stop unit for transmitting a control signal for stopping polishing to the polishing unit when the remaining time estimated by the determination unit has elapsed is further provided.

The end point detecting device according to the fifth aspect of the embodiment is an end point detecting device according to the first or third aspect.
The determination unit inputs measurement data from the start of polishing to the present time output from each of the plurality of types of end point detection sensors at the time of new polishing, and estimates whether or not the present time is the timing of the end point indicating the end of polishing. At the same time, it is estimated and output whether or not the current polishing conditions are normal.

The end point detecting device according to the sixth aspect of the embodiment is an end point detecting device according to a fifth aspect that cites the third aspect.
The first polishing stop unit transmits a control signal for stopping polishing to the polishing unit when it is estimated by the determination unit that the current polishing conditions are normal and the current time is the timing of the end point.

The end point detection device according to the seventh aspect of the embodiment is the end point detection device according to the second or fourth aspect.
The determination unit estimates the time from the present time to the timing of the end point indicating the end of polishing by inputting the measurement data from the start of polishing to the present time output from each of the plurality of types of end point detection sensors at the time of new polishing. , Estimates and outputs whether or not the current polishing conditions are normal.

The end point detection device according to the eighth aspect of the embodiment is an end point detection device according to the seventh aspect that cites the fourth aspect.
The first polishing stop unit transmits a control signal for stopping polishing to the polishing unit when it is estimated by the determination unit that the current polishing conditions are normal and the remaining time is zero.

The end point detecting device according to the ninth aspect of the embodiment is an end point detecting device according to any one of the fifth to eighth aspects.
The determination unit further includes a second polishing stop unit that transmits a control signal for stopping polishing to the polishing unit and issues an alarm when it is determined that the current polishing state is abnormal.

The end point detecting device according to the tenth aspect of the embodiment is an end point detecting device according to any one of the first to ninth aspects.
The plurality of types of end point detection sensors are an optical sensor that shines light on an object to be polished and monitors a change in its reflectance, and a vortex current that applies a line of magnetic force to the object to be polished and monitors a change in the line of magnetic force due to a vortex current generated there. Vibration of the sensor, rocking torque sensor that monitors the change in torque applied to the rocking mechanism that swings the top ring, rotational torque sensor that monitors the change in torque applied to the rotating mechanism that rotates the polishing table, vibration of the top ring or polishing table There are two or more types of vibration sensors that monitor the change in sound generated from the contact portion between the object to be polished and the polishing pad.

The end point detection device according to the eleventh aspect of the embodiment is an end point detection device according to the first aspect.
The trained model has a waveform of measurement data from each of the plurality of types of end point detection sensors output during past polishing from the start of polishing to the end of polishing, and from the start of polishing to the end of polishing acquired during the past polishing. Machine learning the relationship between the polishing pad temperature, slurry temperature, slurry flow rate, pressure in each pressure chamber of the top ring, and one or more auxiliary information of the number of times the polishing pad has been used.
The determination unit receives the measurement data from the start of polishing to the present time output from each of the plurality of types of end point detection sensors at the time of new polishing and the auxiliary information acquired from the start of polishing to the present time, and the present time is It estimates and outputs whether or not it is the timing of the end point.

According to such an aspect, a trained model in which the relationship between the waveform of the measurement data from the start of polishing to the end of polishing and the auxiliary information output from each of the plurality of types of end point detection sensors during past polishing is machine-learned is obtained. By using it, it is possible to estimate and output whether or not the current time is the timing of the end point in consideration of the similarity between the measurement data at the time of past polishing and the auxiliary information. Therefore, it is possible to optimally combine and use the measurement data of a plurality of types of end point detection sensors and auxiliary information, and it is possible to further improve the accuracy of end point detection.

The end point detecting device according to the twelfth aspect of the embodiment is an end point detecting device according to the second aspect.
The trained model has a waveform of measurement data from each of the plurality of types of end point detection sensors output during past polishing from the start of polishing to the end of polishing, and from the start of polishing to the end of polishing acquired during the past polishing. Machine learning the relationship between the polishing pad temperature, slurry temperature, slurry flow rate, pressure in each pressure chamber of the top ring, and one or more auxiliary information of the number of times the polishing pad has been used.
The determination unit receives the measurement data from the start of polishing to the present time output from each of the plurality of types of end point detection sensors at the time of new polishing and the auxiliary information acquired from the start of polishing to the present time as input from the present time. The remaining time until the end point timing indicating the end of polishing is estimated and output.

According to such an aspect, a trained model in which the relationship between the waveform of the measurement data from the start of polishing to the end of polishing and the auxiliary information output from each of the plurality of types of end point detection sensors during past polishing is machine-learned is obtained. By using this, the remaining time from the present time to the timing of the end point indicating the end of polishing can be estimated and output in consideration of the similarity between the measurement data at the time of past polishing and the auxiliary information. Therefore, it is possible to optimally combine and use the measurement data of a plurality of types of end point detection sensors and auxiliary information, and it is possible to further improve the accuracy of end point detection.

The end point detecting device according to the thirteenth aspect of the embodiment is an end point detecting device according to any one of the first to twelfth aspects.
Further, a timing adjusting unit is further provided, in which the timings between the measurement data output from each of the plurality of types of end point detection sensors are matched and then input to the determination unit.

According to such an aspect, it becomes possible to more accurately grasp the waveform of the measurement data output from each of the plurality of types of end point detection sensors, thereby making it possible to determine whether or not the current end point timing is reached. It is possible to make an accurate judgment. Therefore, the accuracy of end point detection can be further improved.

The end point detecting device according to the 14th aspect of the embodiment is an end point detecting device according to the 13th aspect.
The timing adjustment unit simultaneously inputs a timing synchronization signal to the plurality of types of end point detection sensors, and determines the timing of a pulse portion caused by the timing synchronization signal in the measurement data output from each of the plurality of end point detection sensors. By matching, the timings between the measurement data are matched.

The end point detection device according to the fifteenth aspect of the embodiment is an end point detection device according to any one of the first to fourteenth aspects.
The trained model machine-learns the waveform of the measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors provided in the first polishing unit during the past polishing. Machine learning the waveform of the measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors provided in the second polishing unit different from the first polishing unit during the past polishing. There is.

According to such an aspect, it is possible to increase the learning speed of the trained model, which makes it possible to further improve the accuracy of end point detection.

The end point detecting device according to the 16th aspect of the embodiment is an end point detecting device according to the 15th aspect.
The first polishing unit and the second polishing unit are installed in the same factory.

The end point detecting device according to the 17th aspect of the embodiment is an end point detecting device according to the 15th aspect.
The first polishing unit and the second polishing unit are installed in different factories.

The end point detection method according to the eighteenth aspect of the embodiment is
Using a learned model in which the waveforms of the measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors provided in one polishing unit during past polishing are machine-learned, the plurality of types are described. A judgment step is performed in which real-time measurement data from the start of polishing to the present time, which is output from each of the end point detection sensors at the time of new polishing, is input, and whether or not the present time is the timing of the end point indicating the end of polishing is estimated and output. Including.

The end point detection method according to the nineteenth aspect of the embodiment is
Using a learned model in which the waveforms of the measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors provided in one polishing unit during past polishing are machine-learned, the plurality of types are described. It includes a determination step in which the measurement data from the start of polishing to the present time, which is output from each of the end point detection sensors at the time of new polishing, is input, and the remaining time from the present time to the timing of the end point indicating the end of polishing is estimated and output.

The end point detection program according to the twentieth aspect of the embodiment is
Computer,
It has a learned model in which the waveforms of the measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors provided in one polishing unit during past polishing are machine-learned, and the plurality of types are described. The measurement data from the start of polishing to the present time, which is output from each of the end point detection sensors at the time of new polishing, is input, and it is made to function as a judgment unit that estimates and outputs whether or not the present time is the timing of the end point indicating the end of polishing. ..

The end point detection program according to the 21st aspect of the embodiment is
Computer,
It has a learned model in which the waveforms of the measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors provided in one polishing unit during past polishing are machine-learned, and the plurality of types are described. The measurement data from the start of polishing to the present time, which is output from each of the end point detection sensors at the time of new polishing, is input, and the remaining time from the present time to the timing of the end point indicating the end of polishing is estimated and output as a judgment unit.

The computer-readable recording medium according to the 22nd aspect of the embodiment is
Computer,
It has a learned model in which the waveforms of the measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors provided in one polishing unit during past polishing are machine-learned, and the plurality of types are described. The end point that functions as a judgment unit that estimates and outputs whether or not the current time is the timing of the end point indicating the end of polishing by inputting real-time measurement data during polishing output from each of the end point detection sensors during new polishing. The detection program is recorded non-transitory.

The computer-readable recording medium according to the 23rd aspect of the embodiment is
Computer,
It has a learned model in which the waveforms of the measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors provided in one polishing unit during past polishing are machine-learned, and the plurality of types are described. The end point that functions as a judgment unit that estimates and outputs the remaining time from the current point to the timing of the end point indicating the end of polishing by inputting the measurement data from the start of polishing to the present time output from each of the end point detection sensors at the time of new polishing. The detection program is recorded non-transitory.

The trained model (tuned neural network system) according to the 24th aspect of the embodiment is
It has an input layer, one or more intermediate layers connected to the input layer, and an output layer connected to the intermediate layer, and is past from each of a plurality of types of end point detection sensors provided in one polishing unit. The measurement data from the start of polishing to each time point output during polishing is input to the input layer, and the output result output from the output layer is compared with the information on whether or not the time point is the end point timing. By repeating the process of updating the parameters of each node according to the error for the measurement data from the start of polishing in the past polishing to each time point, each of the plurality of types of end point detection sensors outputs during the past polishing. This is a machine-learned waveform of the measurement data from the start of polishing to the end of polishing.
When the measurement data from the start of polishing to the present time, which is output from each of the plurality of types of end point detection sensors at the time of new polishing, is input to the input layer, it is estimated whether or not the present time is the timing of the end point indicating the end of polishing. And make the computer function to output from the output layer.

The trained model (tuned neural network system) according to the 25th aspect of the embodiment is
It has an input layer, one or more intermediate layers connected to the input layer, and an output layer connected to the intermediate layer, and is past from each of a plurality of types of end point detection sensors provided in one polishing unit. The measurement data from the start of polishing to each time point output during polishing is input to the input layer, and the output result output from the output layer and the information on the remaining time from that time point to the timing of the end point indicating the end of polishing By repeating the process of comparing the above and updating the parameters of each node according to the error for the measurement data from the start of polishing to each time point during the past polishing, each of the plurality of types of end point detection sensors in the past This is a machine-learned waveform of the measurement data output from the start of polishing to the end of polishing, which is output during polishing.
When the measurement data from the start of polishing to the present time, which is output from each of the plurality of types of end point detection sensors at the time of new polishing, is input to the input layer, the remaining time from the present time to the timing of the end point indicating the end of polishing is estimated. Make the computer function to output from the output layer.

The specific examples of the embodiments will be described in detail below with reference to the attached drawings. In the following description and the drawings used in the following description, the same reference numerals are used for parts that can be configured in the same manner, and duplicate description is omitted.

<Overall configuration of board processing equipment>
FIG. 1 is a plan view showing the overall configuration of the substrate processing apparatus according to the embodiment. As shown in FIG. 1, this substrate processing apparatus includes a substantially rectangular housing 61. The housing 61 has a side wall 700. The inside of the housing 61 is divided into a load / unload portion 62, a polishing portion 63, and a cleaning portion 64 by partition walls 1a and 1b. The load / unload section 62, the polishing section 63, and the cleaning section 64 are assembled independently and exhausted independently. Further, the substrate processing apparatus has a control unit 65 that controls the substrate processing operation.

The load / unload section 62 includes two or more (four in this embodiment) front load sections 20 on which wafer cassettes for stocking a large number of wafers (boards) are placed. These front load portions 20 are arranged adjacent to the housing 61, and are arranged along the width direction (direction perpendicular to the longitudinal direction) of the substrate processing apparatus. The front load unit 20 can be equipped with an open cassette, a SMIF (Standard Manufacturing Interface) pod, or a FOUP (Front Opening Unified Pod). Here, SMIF and FOUP are closed containers that can maintain an environment independent of the external space by storing the wafer cassette inside and covering it with a partition wall.

Further, in the load / unload section 62, a traveling mechanism 21 is laid along the line of the front load section 20. Two transfer robots (loaders) 22 that can move along the arrangement direction of the wafer cassettes are installed on the traveling mechanism 21. The transfer robot 22 can access the wafer cassette mounted on the front load unit 20 by moving on the traveling mechanism 21. Each transfer robot 22 has two hands on the top and bottom. The upper hand is used to return the processed wafer to the wafer cassette. The lower hand is used to remove the unprocessed wafer from the wafer cassette. In this way, the upper and lower hands can be used properly. Further, the lower hand of the transfer robot 22 can invert the wafer by rotating around its axis.

The load / unload section 62 is an area that needs to be kept in the cleanest state. Therefore, the inside of the load / unload section 62 is always maintained at a pressure higher than that of the outside of the substrate processing device, the polishing section 63, and the cleaning section 64. The polishing unit 63 is the dirtiest region because a slurry is used as the polishing liquid. Therefore, a negative pressure is formed inside the polishing portion 63, and the pressure is maintained lower than the internal pressure of the cleaning portion 64. The load / unload section 62 is provided with a filter fan unit (not shown) having a clean air filter such as a HEPA filter, a ULPA filter, or a chemical filter. Clean air from which particles, toxic steam, and toxic gas have been removed is constantly blown out from the filter fan unit.

The polishing unit 63 is an area where the wafer is polished (flattened), and includes a first polishing unit 3A, a second polishing unit 3B, a third polishing unit 3C, and a fourth polishing unit 3D. As shown in FIG. 1, the first polishing unit 3A, the second polishing unit 3B, the third polishing unit 3C, and the fourth polishing unit 3D are arranged along the longitudinal direction of the substrate processing apparatus.

As shown in FIG. 1, the first polishing unit 3A includes a polishing table 30A, a top ring 31A, a polishing liquid supply nozzle 32A, a dresser 33A, and an atomizer 34A. A polishing pad 10 having a polishing surface is attached to the polishing table 30A. The top ring (holding portion) 31A holds the wafer and polishes the wafer while pressing it against the polishing pad 10 on the polishing table 30A. The polishing liquid supply nozzle 32A supplies a polishing liquid or a dressing liquid (for example, pure water) to the polishing pad 10. The dresser 33A dresses the polished surface of the polishing pad 10. The atomizer 34A atomizes a mixed fluid of a liquid (for example, pure water) and a gas (for example, nitrogen gas) or a liquid (for example, pure water) and injects it onto the polished surface.

Similarly, the second polishing unit 3B includes a polishing table 30B to which the polishing pad 10 is attached, a top ring 31B, a polishing liquid supply nozzle 32B, a dresser 33B, and an atomizer 34B. The third polishing unit 3C includes a polishing table 30C to which a polishing pad 10 is attached, a top ring 31C, a polishing liquid supply nozzle 32C, a dresser 33C, and an atomizer 34C. The fourth polishing unit 3D includes a polishing table 30D to which a polishing pad 10 is attached, a top ring 31D, a polishing liquid supply nozzle 32D, a dresser 33D, and an atomizer 34D.

Since the first polishing unit 3A, the second polishing unit 3B, the third polishing unit 3C, and the fourth polishing unit 3D have the same configuration as each other, the details of the polishing unit will be described below in the first polishing. The unit 3A will be described as a target.

<Structure of polishing unit>
FIG. 2 is a perspective view schematically showing the first polishing unit 3A. The top ring 31A is supported by the top ring shaft 636. A polishing pad 10 is attached to the upper surface of the polishing table 30A, and the upper surface of the polishing pad 10 constitutes a polishing surface for polishing the semiconductor wafer 16. It should be noted that fixed abrasive grains can be used instead of the polishing pad 10. The top ring 31A and the polishing table 30A are configured to rotate about their axis, as indicated by the arrows. The semiconductor wafer 16 is held on the lower surface of the top ring 31A by vacuum suction. At the time of polishing, the polishing liquid is supplied from the polishing liquid supply nozzle 32A to the polishing surface of the polishing pad 10, and the semiconductor wafer 16 to be polished is pressed against the polishing surface by the top ring 31A to be polished.

FIG. 3 is a cross-sectional view schematically showing the structure of the top ring 31A. The top ring 31A is connected to the lower end of the top ring shaft 636 via a universal joint 637. The universal joint 637 is a ball joint that transmits the rotation of the top ring shaft 636 to the top ring 31A while allowing the top ring 31A and the top ring shaft 636 to tilt each other. The top ring 31A includes a substantially disk-shaped top ring main body 638 and a retainer ring 640 arranged below the top ring main body 638. The top ring body 638 is formed of a material having high strength and rigidity such as metal and ceramics. Further, the retainer ring 640 is formed of a highly rigid resin material, ceramics, or the like. The retainer ring 640 may be integrally formed with the top ring main body 638.

In the space formed inside the top ring main body 638 and the retainer ring 640, a circular elastic pad 642 that abuts on the semiconductor wafer 16, an annular pressure sheet 643 made of an elastic film, and an elastic pad 642 are held. A roughly disk-shaped chucking plate 644 is housed. The upper peripheral end of the elastic pad 642 is held by the chucking plate 644, and four pressure chambers (airbags) P1, P2, P3, P4 are provided between the elastic pad 642 and the chucking plate 644. There is. The pressure chambers P1, P2, P3, and P4 are formed by an elastic pad 642 and a chucking plate 644. Pressurized fluid such as pressurized air is supplied to the pressure chambers P1, P2, P3, and P4 via the fluid passages 651,652,653,654, respectively, or is evacuated. The central pressure chamber P1 is circular, and the other pressure chambers P2, P3, and P4 are annular. These pressure chambers P1, P2, P3 and P4 are concentrically arranged.

The internal pressures of the pressure chambers P1, P2, P3, and P4 can be changed independently of each other by the pressure adjusting unit described later, whereby the four regions of the semiconductor wafer 16, that is, the central portion and the inner intermediate portion, can be changed. , The pressing force on the outer middle part, and the peripheral part can be adjusted independently. Further, by raising and lowering the entire top ring 31A, the retainer ring 640 can be pressed against the polishing pad 10 with a predetermined pressing force. A pressure chamber P5 is formed between the chucking plate 644 and the top ring main body 638, and a pressurized fluid is supplied to the pressure chamber P5 via a fluid passage 655 or is evacuated. There is. As a result, the entire chucking plate 644 and the elastic pad 642 can move in the vertical direction.

The peripheral end of the semiconductor wafer 16 is surrounded by a retainer ring 640 so that the semiconductor wafer 16 does not pop out from the top ring 31A during polishing. An opening (not shown) is formed in a portion of the elastic pad 642 constituting the pressure chamber P3 so that the semiconductor wafer 16 is adsorbed and held by the top ring 31A by forming a vacuum in the pressure chamber P3. It has become. Further, the semiconductor wafer 16 is released from the top ring 31A by supplying nitrogen gas, dry air, compressed air, or the like to the pressure chamber P3.

FIG. 4 is a cross-sectional view schematically showing another structural example of the top ring 31A. In this example, the chucking plate is not provided, and the elastic pad 642 is attached to the lower surface of the top ring body 638. Further, the pressure chamber P5 between the chucking plate and the top ring main body 638 is not provided. Instead of this, an elastic bag 646 is arranged between the retainer ring 640 and the top ring main body 638, and a pressure chamber P6 is formed inside the elastic bag 646. The retainer ring 640 can move up and down relative to the top ring body 638. A fluid passage 656 communicates with the pressure chamber P6, and a pressurized fluid such as pressurized air is supplied to the pressure chamber P6 through the fluid passage 656. The internal pressure of the pressure chamber P6 can be adjusted by a pressure adjusting unit described later. Therefore, the pressing force on the polishing pad 10 of the retainer ring 640 can be adjusted independently of the pressing force on the semiconductor wafer 16. Other configurations and operations are the same as the configuration of the top ring shown in FIG. In this embodiment, either type of top ring of FIG. 3 or FIG. 4 can be used.

FIG. 5 is a cross-sectional view for explaining a mechanism for rotating and swinging the top ring 31A. The top ring shaft (eg, spline shaft) 636 is rotatably supported by the top ring head 660. Further, the top ring shaft 636 is connected to the rotating shaft of the motor M1 via

pulleys

661 and 662 and the belt 663, and the motor M1 rotates the top ring shaft 636 and the top ring 31A around the axis thereof. The motor M1 is attached to the upper part of the top ring head 660. Further, the top ring head 660 and the top ring shaft 636 are connected by an air cylinder 665 as a vertical drive source. The air (compressed gas) supplied to the air cylinder 665 causes the top ring shaft 636 and the top ring 31A to move up and down integrally. Instead of the air cylinder 665, a mechanism having a ball screw and a servomotor may be used as a vertical drive source.

The top ring head 660 is rotatably supported by a support shaft 667 via a bearing 672. The support shaft 667 is a fixed shaft and has a structure that does not rotate. A motor M2 is installed on the top ring head 660, and the relative positions of the top ring head 660 and the motor M2 are fixed. The rotation shaft of the motor M2 is connected to the support shaft 667 via a rotation transmission mechanism (gear or the like) (not shown), and by rotating the motor M2, the top ring head 660 swings around the support shaft 667. It is designed to (swing). Therefore, due to the swinging motion of the top ring head 660, the top ring 31A supported at the tip thereof moves between the polishing position above the polishing table 30A and the lateral transport position of the polishing table 30A. In the present embodiment, the swing mechanism for swinging the top ring 31A is composed of the motor M2. As shown in FIG. 5, a swing torque sensor 26 for detecting torque applied to the swing mechanism is connected to the swing mechanism (motor M2) that swings the top ring 31A. The signal of the oscillating torque sensor 26 is transmitted to the control unit 65, which will be described later.

Inside the top ring shaft 36, a through hole (not shown) extending in the longitudinal direction thereof is formed. The

fluid passages

651, 652, 652, 654, 655, 656 of the top ring 31A described above are connected to the rotary joint 669 provided at the upper end of the top ring shaft 636 through the through holes. A fluid such as pressurized gas (clean air) or nitrogen gas is supplied to the top ring 31A via the rotary joint 669, and the gas is evacuated from the top ring 31A. A plurality of fluid pipes 670 communicating with the fluid passages 651,652,655,654,655,656 (see FIGS. 3 and 4) are connected to the rotary joint 669, and these fluid pipes 670 are connected to the pressure adjusting unit 675. It is connected. Further, a fluid pipe 671 that supplies pressurized air to the air cylinder 665 is also connected to the pressure adjusting unit 675.

The pressure adjusting unit 675 includes an electropneumatic regulator that adjusts the pressure of the fluid supplied to the top ring 31A, pipes connected to the

fluid pipes

670 and 671, air operated valves provided in these pipes, and these air operated valves. It has an electropneumatic regulator that adjusts the pressure of air that is the operating source of the above, an ejector that forms a vacuum on the top ring 31A, and the like, and these are collectively formed as one block (unit). The pressure adjusting unit 675 is fixed to the upper part of the top ring head 660. The pressure of the pressurized gas supplied to the pressure chambers P1, P2, P3, P4, P5 (see FIG. 3) of the top ring 31A and the pressurized air supplied to the air cylinder 665 is the electric pressure of the pressure adjusting unit 675. Adjusted by an empty regulator. Similarly, a vacuum is formed in the airbags P1, P2, P3, P4 of the top ring 31A and in the pressure chamber P5 between the chucking plate 44 and the top ring main body 38 by the ejector of the pressure adjusting unit 675.

In this way, since the electropneumatic regulator and the valve, which are pressure adjusting devices, are installed near the top ring 31A, the controllability of the pressure in the top ring 31A is improved. More specifically, since the distance between the electropneumatic regulator and the pressure chambers P1, P2, P3, P4, and P5 is short, the responsiveness to the pressure change command from the control unit 65 is improved. Similarly, since the ejector, which is a vacuum source, is also installed near the top ring 31A, the responsiveness when forming a vacuum in the top ring 31A is improved. Further, the back surface of the pressure adjusting unit 675 can be used as a mounting pedestal for electrical equipment, and a mounting frame, which has been conventionally required, can be eliminated.

The top ring head 660, top ring 31A, pressure adjusting unit 675, top ring shaft 636, motor M1, motor M2, and air cylinder 665 are configured as one module (hereinafter referred to as top ring assembly). That is, the top ring shaft 636, the motor M1, the motor M2, the pressure adjusting unit 675, and the air cylinder 665 are attached to the top ring head 660. The top ring head 660 is configured to be removable from the support shaft 667. Therefore, the top ring assembly can be removed from the substrate processing apparatus by separating the top ring head 660 and the support shaft 667. According to such a configuration, the maintainability of the support shaft 667 and the top ring head 660 can be improved. For example, when an abnormal noise is generated from the bearing 672, the bearing 672 can be easily replaced, and when the motor M2 or the rotation transmission mechanism (reducer) is replaced, it is not necessary to remove the adjacent device. ..

FIG. 6 is a cross-sectional view schematically showing the internal structure of the polishing table 30A. As shown in FIG. 6, the polishing table 30A is provided with a rotation mechanism (motor 300) for rotationally driving the polishing table 30A. The power of the motor 300 is transmitted to the spindle 320 fixed to the polishing table 310 via the belt 310 to rotate the polishing table 30A. As shown in FIG. 6, a rotation torque sensor 330 that detects torque applied to the rotation mechanism is connected to a rotation mechanism (motor 300) that rotates the polishing table 30A. The signal of the rotational torque sensor 330 is transmitted to the control unit 65, which will be described later.

As shown in FIG. 6, an eddy current sensor 676A for detecting the state of the film of the semiconductor wafer 16 is embedded inside the polishing table 30A. The signal of the eddy current sensor 676A is transmitted to the control unit 65, and the control unit 65 generates a monitoring signal indicating the film thickness. The value of this monitoring signal (and sensor signal) does not indicate the film thickness itself, but the value of the monitoring signal changes according to the film thickness. Therefore, the monitoring signal can be said to be a signal indicating the film thickness of the semiconductor wafer 16.

The control unit 65 determines the internal pressure of each pressure chamber P1, P2, P3, P4 based on the monitoring signal so that the determined internal pressure is formed in each pressure chamber P1, P2, P3, P4. It is designed to issue a command to the pressure adjusting unit 675. As shown in FIG. 6, the control unit 65 includes a pressure control unit 200 that operates the internal pressure of each pressure chamber P1, P2, P3, P4 based on a monitoring signal, and an end point detection unit 100 that detects the polishing end point. have.

The eddy current sensor 676A is provided on the polishing table of the second polishing unit 3B, the third polishing unit 3C, and the fourth polishing unit 3D as well as the first polishing unit 3A. The control unit 65 generates a monitoring signal from the signal sent from the eddy current sensor 676A of each polishing unit 3A to 3D, and monitors the progress of wafer polishing in each polishing unit 3A to 3D. When a plurality of wafers are polished by the polishing units 3A to 3D, the control unit 5 monitors a monitoring signal indicating the thickness of the wafer during polishing, and based on the monitoring signals, the polishing units 3A to 3D monitor the monitoring signals. The pressing force of the top rings 31A to 31D is controlled so that the polishing times of the top rings are substantially the same. By adjusting the pressing force of the top rings 31A to 31D during polishing based on the monitoring signal in this way, the polishing time in the polishing units 3A to 3D can be leveled.

The semiconductor wafer 16 may be polished by any of the first polishing unit 3A, the second polishing unit 3B, the third polishing unit 3C, and the fourth polishing unit 3D, or is selected in advance from these polishing units 3A to 3D. It may be continuously polished by a plurality of polishing units. For example, the semiconductor wafer 16 may be polished in the order of the first polishing unit 3A → the second polishing unit 3B, or the semiconductor wafer 16 may be polished in the order of the third polishing unit 3C → the fourth polishing unit 3D. .. Further, the semiconductor wafer 16 may be polished in the order of the first polishing unit 3A → the second polishing unit 3B → the third polishing unit 3C → the fourth polishing unit 3D. In any case, the throughput can be improved by leveling all the polishing times of the polishing units 3A to 3D.

The eddy current sensor 676A is preferably used when the wafer film is a metal film. When the film of the wafer is a film having light transmittance such as an oxide film, an optical sensor may be used instead of the eddy current sensor 676A or together with the eddy current sensor 676A. Alternatively, a microwave sensor may be used in place of the eddy current sensor 676A or in combination with the eddy current sensor 676A. The microwave sensor can be used for both metal and non-metal films. Hereinafter, an example of an optical sensor and a microwave sensor will be described.

FIG. 7 is a schematic view for explaining the optical sensor 676B provided on the polishing table 30A. As shown in FIG. 7, an optical sensor 676B for detecting the state of the film of the semiconductor wafer 16 is embedded inside the polishing table 30A. The optical sensor 676B irradiates the semiconductor wafer 16 with light and detects the state of the film (thickness, etc.) of the semiconductor wafer 16 from the intensity (reflection intensity or reflectance) of the reflected light from the semiconductor wafer 16.

Further, the polishing pad 10 is provided with a light transmitting portion 677 for transmitting light from the optical sensor 676B. The translucent portion 677 is made of a material having a high transmittance, and is formed of, for example, quartz glass, a glass material, pure water (with a flow path), or the like. Alternatively, the light-transmitting portion 677 may be formed by providing a through hole in the polishing pad 10 and allowing a transparent liquid to flow from below while the through hole is closed by the semiconductor wafer 16. The translucent portion 677 is arranged at a position where it passes through the center of the semiconductor wafer 16 held by the top ring 31A.

As shown in FIG. 7, the optical sensor 676B includes a light source 678a, a light emitting optical fiber 678b as a light emitting portion that irradiates the surface to be polished of the semiconductor wafer 16 with light from the light source 678a, and reflected light from the surface to be polished. A light receiving optical fiber 678c as a light receiving unit that receives light, a spectroscope that disperses the light received by the light receiving optical fiber 678c, and a plurality of light receiving elements that store the light dispersed by the spectroscope as electrical information are inside. The spectroscope unit 678d, the operation control unit 678e that controls the on / off of the light source 678a and the timing of reading start of the light receiving element in the spectroscope unit 678d, and the power supply 678f that supplies power to the operation control unit 678e. And have. Power is supplied to the light source 678a and the spectroscope unit 678d via the operation control unit 678e.

The light emitting end of the light emitting optical fiber 678b and the light receiving end of the light receiving optical fiber 678c are configured to be substantially perpendicular to the surface to be polished of the semiconductor wafer 16. As the light receiving element in the spectroscope unit 678d, for example, a 128-element photodiode array can be used. The spectroscope unit 678d is connected to the operation control unit 678e. Information from the light receiving element in the spectroscope unit 678d is sent to the operation control unit 678e, and spectral data of reflected light is generated based on this information. That is, the motion control unit 678e reads the electrical information stored in the light receiving element and generates the spectrum data of the reflected light. This spectral data shows the intensity of the reflected light decomposed according to the wavelength, and changes depending on the film thickness.

The operation control unit 678e is connected to the control unit 65 described above. In this way, the spectrum data generated by the operation control unit 678e is transmitted to the control unit 65. The control unit 65 calculates a characteristic value associated with the film thickness of the semiconductor wafer 16 based on the spectrum data received from the operation control unit 678e, and uses this as a monitoring signal.

FIG. 8 is a schematic diagram for explaining the microwave sensor 676C provided on the polishing table 30A. The microwave sensor 676C is a waveguide that connects an antenna 680a that irradiates the surface to be polished of the semiconductor wafer 16 with microwaves, a sensor body 680b that supplies microwaves to the antenna 680a, and the antenna 680a and the sensor body 680b. It is equipped with a tube 681. The antenna 680a is embedded in the polishing table 30A and is arranged so as to face the center position of the semiconductor wafer 16 held by the top ring 31A. The antenna 680a may be located anywhere on the trajectory through which the center of the polishing head 31A passes when the polishing head (top ring) 31A is swung.

The sensor body 680b includes a microwave source 680c that generates microwaves and supplies microwaves to the antenna 680a, microwaves (incident waves) generated by the microwave source 680c, and microwaves (reflection) reflected from the surface of the semiconductor wafer 16. It is provided with a separator 680d for separating the wave) and a detection unit 680e for receiving the reflected wave separated by the separator 680d and detecting the amplitude and phase of the reflected wave. As the separator 680d, a directional coupler is preferably used.

The antenna 680a is connected to the separator 680d via a waveguide 681. The microwave source 680c is connected to the separator 680d, and the microwave generated by the microwave source 680c is supplied to the antenna 680a via the separator 680d and the waveguide 681. The microwave is irradiated from the antenna 680a toward the semiconductor wafer 16 and passes through (penetrates) the polishing pad 610 to reach the semiconductor wafer 16. The reflected wave from the semiconductor wafer 16 passes through the polishing pad 10 again and is received by the antenna 680a.

The reflected wave is sent from the antenna 680a to the separator 680d via the waveguide 681, and the incident wave and the reflected wave are separated by the separator 680d. The reflected wave separated by the separator 680d is transmitted to the detection unit 680e. The detection unit 680e detects the amplitude and phase of the reflected wave. The amplitude of the reflected wave is detected as electric power (dbm or W) or voltage (V), and the phase of the reflected wave is detected by a phase measuring instrument (not shown) built in the detection unit 680e. The amplitude and phase of the reflected wave detected by the detection unit 680e are sent to the control unit 65, where the film thickness of the metal film or non-metal film of the semiconductor wafer 16 is analyzed based on the amplitude and phase of the reflected wave. .. The analyzed value is monitored by the control unit 65 as a monitoring signal.

<Configuration of end point detector>
Next, the configuration of the end point detection unit 100 (end point detection device) included in the control unit 65 described above will be described. FIG. 9 is a block diagram showing the configuration of the end point detection unit 100.

As shown in FIG. 9, the end point detecting unit 100 includes a timing adjusting unit 110, a determination unit 120, a first polishing stop unit 130, and a second polishing stop unit 140.

The determination unit 120 polishes from the start of polishing output from each of a plurality of types of end point detection sensors (first to third sensors 51 to 53 in the illustrated example) provided in one polishing unit 3A during past polishing. It has a trained model 121 (for example, a tuned neural network system) in which the waveform of the measurement data up to the end is machine-learned.

The plurality of types of end point detection sensors (first to third sensors 51 to 53) are an optical sensor 676B (see FIG. 7) that irradiates light on the object to be polished and monitors the change in the reflectance, and magnetic lines of force on the object to be polished. A vortex current sensor 676A (see FIG. 6) that monitors changes in magnetic field lines due to eddy currents generated there, and a swing torque sensor that monitors changes in torque applied to the swing mechanism (motor M2) that swings the top ring 31A. 26 (see FIG. 5), rotational torque sensor 330 (see FIG. 6) that monitors changes in torque applied to the rotating mechanism (motor 300) that rotates the polishing table 30A, vibration that monitors the vibration of the top ring 31A or polishing table 30A. Two or more of a sensor (not shown) and a sound sensor (not shown) that monitors a change in sound generated from a contact portion between the wafer 16 and the polishing pad 10.

The learning method (neural network system tuning method) of the trained model 121 may be supervised learning, unsupervised learning, or reinforcement learning. FIG. 10 is a schematic diagram for explaining an example of the configuration of the trained model 121. As shown in FIG. 10, the trained model 121 is a hierarchical neural network or quantum neural network having an input layer, one or more intermediate layers connected to the input layer, and an output layer connected to the intermediate layer. It may include a network (QNN). The trained model 121 may include a neural network in which intermediate layers are multi-layered, that is, deep learning (deep learning).

As one form, an example of a method of generating the trained model 121 will be described. As shown in FIG. 10A, a plurality of types of end point detection sensors (first to third sensors 51 to 53) provided in one polishing unit 3A will be described. ), The measurement data from the start of polishing to each time point output during the past polishing is input to the input layer, and the output result output from the output layer is linked to the measurement data at that time point. Compare the information on whether the polishing conditions at the time point are normal and whether the time point is the timing of the end point (that is, the end of polishing), and the parameters (weights and thresholds) of each node according to the error. The process of updating) is repeated for the measurement data from the start of polishing to each time point in the past polishing. As a result, the waveform of the measurement data from the start to the end of polishing output from the plurality of types of end point detection sensors (first to third sensors 51 to 53) provided in one polishing unit 3A during the past polishing is machined. The trained trained model 121 (tuned neural network system) is generated.

As another embodiment, an example of a method of generating the trained model 121 will be described. As shown in FIG. 10B, a plurality of types of end point detection sensors (first to third sensors 51 to 53) provided in one polishing unit 3A will be described. ), The measurement data from the start of polishing to each time point output during the past polishing is input to the input layer, and the output result output from the output layer is linked to the measurement data at that time point. Compare the information on whether the polishing conditions at that point in time are normal and the remaining time from that point in time to the end point (that is, the end of polishing), and set the parameters (weight, threshold, etc.) of each node according to the error. The process of updating is repeated for the measurement data from the start of polishing to each time point in the past polishing. As a result, the waveform of the measurement data from the start to the end of polishing output from the plurality of types of end point detection sensors (first to third sensors 51 to 53) provided in one polishing unit 3A during the past polishing is machined. The trained trained model 121 (tuned neural network system) is generated.

Regarding the measurement data (teacher data) to be learned by the trained model 121, the past from each of the plurality of types of end point detection sensors (first to third sensors 51 to 53) provided in one polishing unit 3A. Of the measurement data from the start of polishing to the end of polishing output during polishing, the measurement data when the polishing is completed normally without over-polishing or insufficient polishing is for each time point from the start of polishing to the end of polishing. It may be associated that all the polishing conditions were normal.

FIG. 12 is an image diagram for explaining the learning content of the trained model 121. In the example shown in FIG. 12, the regions A1 to A3 are each in the past polishing time output from each of the first to third sensors 51 to 53 when the polishing conditions from the start to the end of polishing are normal. The waveform of the measurement data of the above indicates a region that is statistically included with a predetermined probability (for example, a reliability CL of 95% or more). In the example shown in FIG. 12, since the width of the region A1 is narrower than the widths of the regions A2 and A3 in the time zone close to the polishing start time, the output is output from the first sensor 51 in the time zone close to the polishing start time. It can be interpreted that the priority of the measurement data to be measured is higher than the priority of the measurement data output from the second and

third sensors

52 and 53. On the other hand, since the width of the region A2 is narrower than the widths of the regions A1 and A3 in the time zone near the end of polishing, the measurement data output from the second sensor 52 is in the time zone near the end of polishing. It can be interpreted that the priority is higher than the priority of the measurement data output from the first and

third sensors

51 and 53.

The trained model 121 machine-learns the waveform of the measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors provided in the first polishing unit 3A during the past polishing. , Machine learning of the waveform of the measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors provided in the second polishing unit 3B, which is different from the first polishing unit 3A, during the past polishing. You may be doing it. In this case, the learning speed of the trained model 121 can be increased. Here, as shown in FIG. 1, the first polishing unit 3A and the second polishing unit 3B may be installed in the same factory. Alternatively, although not shown, the first polishing unit 3A and the second polishing unit 3B may be installed in different factories.

As one form, as shown in FIG. 10A, the determination unit 120 is used for new polishing from each of a plurality of types of end point detection sensors (first to third sensors 51 to 53) provided in one polishing unit 3A. Using the output measurement data from the start of polishing to the present time as input, using the trained model 121 according to the above example, whether or not the current polishing conditions are normal, and the timing of the end point indicating the end of polishing at the present time. It is estimated and output whether or not it is. As another form, as shown in FIG. 10B, the determination unit 120 is used during new polishing from each of the plurality of types of end point detection sensors (first to third sensors 51 to 53) provided in one polishing unit 3A. Using the output measurement data from the start of polishing to the present time as input, the above-mentioned trained model 121 is used to estimate whether or not the current polishing conditions are normal, and the timing of the end point indicating the end of polishing from the present time. The remaining time until is estimated and output.

For example, as shown in FIG. 13, the measurement data D1 to D3 from the start of polishing to the present time output from each of the first to third sensors 51 to 53 at the time of new polishing are from the start of polishing to the end of polishing, respectively. If the area A1 to A3 that should be statistically included when the polishing conditions are normal and the current time does not match the timing of the end of polishing, the determination unit 120 is used as one form. Is output by assuming that the polishing conditions at the present time are normal and the current time is not the timing of the end point indicating the end of polishing. As another form, the determination unit 120 may estimate that the polishing conditions at the present time are normal, and estimate and output the remaining time Te from the present time to the timing of the end point indicating the end of polishing.

Further, for example, as shown in FIG. 14, the measurement data D1 to D3 from the start of polishing to the present time output from each of the first to third sensors 51 to 53 at the time of new polishing are obtained from the start of polishing to the end of polishing, respectively. If it is within the areas A1 to A3 that should be statistically included when the polishing conditions up to are normal, and the current time coincides with the timing of the end of polishing, it is determined as one form. The unit 120 estimates that the current polishing conditions are normal and the current time is the timing of the end point indicating the end of polishing, and outputs the data. As another form, the determination unit 120 may estimate that the polishing conditions at the present time are normal, and estimate that the remaining time Te from the present time to the timing of the end point indicating the end of polishing is zero and output. ..

Further, for example, as shown in FIG. 15, any one of the measurement data D1 to D3 from the start of polishing to the present time output from each of the first to third sensors 51 to 53 at the time of new polishing (in the illustrated example). If the measurement data D1) of the first sensor 51 deviates from the region A1 that should be statistically included when the polishing conditions from the start to the end of polishing are normal, the determination unit 120 is at the present time. It is estimated that the polishing conditions of the above are abnormal and output. It should be noted that FIGS. 12 to 15 are merely "image diagrams" and are not limited to these, and are normal using a trained model trained using measurement data of normal termination without using statistical processing. -A method of outputting an abnormality judgment may be used. For example, a trained model that predicts waveform analysis is constructed (tuned) by learning, and as one form, the determination unit 120 may output the probability *% of the current end point detection as the result of the prediction. .. As another form, the determination unit 120 may output, as a result of the prediction, that the end point has a probability of *% or more after * s of the end point detection.

When the determination unit 120 preferentially uses the measurement data of which end point detection sensor with respect to the measurement data from the start of polishing to the present time output from each of the first to third sensors 51 to 53 at the time of new polishing. By using the above-mentioned trained model 121 (see FIG. 10A) as one form, even if the operator does not explicitly instruct whether to switch the priority order between the end point detection sensors at the timing, the past In view of the similarity with the waveform of the measurement data at the time of polishing, it is possible to estimate and output whether or not the current polishing condition is normal and whether or not the current time is the timing of the end point. Further, as another form, by using the above-mentioned trained model 121 (see FIG. 10B), is the current polishing condition normal in view of the similarity with the waveform of the measurement data at the time of past polishing? Whether or not, and the remaining time from the current time to the end point timing can be estimated and output.

As a modification, the trained model 121 starts from the polishing start output from each of the plurality of types of end point detection sensors (for example, the first to third sensors 51 to 53) provided in one polishing unit 3A during the past polishing. The waveform of the measurement data until the end of polishing, the temperature of the polishing pad 10 from the start of polishing to the end of polishing acquired during the past polishing, the temperature of the slurry, the flow rate of the slurry, the pressure in each pressure chamber of the top ring 31A, and the polishing. Machine learning may be performed on the relationship with one or more auxiliary information of the number of times the pad 10 is used.

As one embodiment, an example of a method of generating the trained model 121 according to one modification will be described. As shown in FIG. 11A, a plurality of types of end point detection sensors (first to first) provided in one polishing unit 3A. The measurement data from the start of polishing to each time point output from each of the 3 sensors 51 to 53) during the past polishing, and the temperature of the polishing pad 10, the temperature of the slurry, and the flow rate of the slurry acquired from the start of polishing to that time point. , The pressure of each pressure chamber of the top ring 31A, and one or more auxiliary information of the number of times the polishing pad 10 has been used are input to the input layer, and the output result output from the output layer and the measurement at that time point. Compare the information linked to the data whether the polishing conditions at that point in time are normal and whether the point in time is the timing of the end point (that is, the end of polishing), and each according to the error. The process of updating node parameters (weights, thresholds, etc.) is repeated for the measurement data from the start of polishing to each time point during the past polishing. As a result, the waveforms of the measurement data from the start of polishing to the end of polishing output from the plurality of types of end point detection sensors (first to third sensors 51 to 53) provided in one polishing unit 3A during the past polishing are obtained. A learned model 121 is generated in which the relationship with the auxiliary information from the start of polishing to the end of polishing acquired at the time of the past polishing is machine-learned.

In this case, as shown in FIG. 11A, the determination unit 120 is output from each of the plurality of types of end point detection sensors (first to third sensors 51 to 53) provided in one polishing unit 3A at the time of new polishing. Whether or not the current polishing conditions are normal using the learned model 121 according to one modification by inputting the measurement data from the start of polishing to the present time and the auxiliary information acquired from the start of polishing to the present time. Or, and whether or not the current time is the timing of the end point is estimated and output.

As another embodiment, an example of a method of generating the trained model 121 according to one modification will be described. As shown in FIG. 11B, a plurality of types of end point detection sensors (first to first) provided in one polishing unit 3A. The measurement data from the start of polishing to each time point output from each of the 3 sensors 51 to 53) during the past polishing, and the temperature of the polishing pad 10, the temperature of the slurry, and the flow rate of the slurry acquired from the start of polishing to that time point. , The pressure of each pressure chamber of the top ring 31A, and one or more auxiliary information of the number of times the polishing pad 10 has been used are input to the input layer, and the output result output from the output layer and the measurement at that time point. Compare the information on whether the polishing conditions at that time point are normal and the remaining time until the end point (that is, the end of polishing) at that point in time, and the parameters of each node (that is, the end of polishing) according to the error. The process of updating (weight, threshold, etc.) is repeated for the measurement data from the start of polishing to each time point in the past polishing. As a result, the waveforms of the measurement data from the start of polishing to the end of polishing output from the plurality of types of end point detection sensors (first to third sensors 51 to 53) provided in one polishing unit 3A during the past polishing are obtained. A learned model 121 is generated in which the relationship with the auxiliary information from the start of polishing to the end of polishing acquired at the time of the past polishing is machine-learned.

In this case, as shown in FIG. 11B, the determination unit 120 is output from each of the plurality of types of end point detection sensors (first to third sensors 51 to 53) provided in one polishing unit 3A at the time of new polishing. Whether or not the current polishing conditions are normal using the learned model 121 according to one modification by inputting the measurement data from the start of polishing to the present time and the auxiliary information acquired from the start of polishing to the present time. Is estimated, and the remaining time from the current time to the end point timing is estimated and output.

Here, the determination unit 120 provides one or more auxiliary information of the temperature of the polishing pad 10, the temperature of the slurry, the flow rate of the slurry, the pressure in each pressure chamber of the top ring 31A, and the number of times the polishing pad 10 is used. It may be obtained from an engineering system (Equipment Engineering System; EES).

With reference to FIG. 9, in one embodiment, the first polishing stop unit 130 performs polishing when the determination unit 120 determines that the current polishing conditions are normal and the current time is the timing of the end point. A control signal for stopping (stopping the operation of the polishing unit 3A) is transmitted to the polishing unit 3A. In another embodiment, the first polishing stop unit 130 may transmit a control signal for stopping polishing to the polishing unit 3A when the remaining time estimated by the determination unit 120 has elapsed.

When it is determined that the current polishing state is abnormal, the second polishing stop unit 140 transmits a control signal for stopping polishing (stopping the operation of the polishing unit 3A) to the polishing unit 3A and issues an alarm. Report. Instead of issuing an alarm, the second polishing stop unit 140 may adopt one method of error display, patrol display, and automatic contact, or two of alarm, error display, patrol display, and automatic contact. A method based on the above combination may be adopted.

The timing adjusting unit 110 matches the timing between the measurement data D1 to D3 output from each of the plurality of types of end point detection sensors (for example, the first to third sensors 51 to 53) provided in one polishing unit 3A. Is input to the determination unit 120.

Specifically, for example, the timing adjusting unit 110 simultaneously inputs timing synchronization signals to a plurality of types of end point detection sensors (first to third sensors 51 to 53) before starting polishing, and as shown in FIG. , Of the measurement data D1 to D3 output from each of the plurality of end point detection sensors (first to third sensors 51 to 53), the timing of the pulse portion caused by the timing synchronization signal is compared. In the example shown in FIG. 16, the pulse portion of the measurement data D2 output from the second sensor 52 and the pulse portion of the measurement data D3 output from the third sensor 53 have the same timing, but are compared with them. , The pulse portion of the measurement data D1 output from the first sensor 51 is advanced (earlier) by Δt.

If there is a deviation in the timing of the pulse portion of the measurement data D1 to D3 output from each of the plurality of end point detection sensors (first to third sensors 51 to 53), the timing adjustment unit 110 may use the pulse portion of the pulse portion. By matching the timings, the timings between the measurement data D1 to D3 are matched. For example, as shown in FIG. 16, when the pulse portion of the measurement data D1 output from the first sensor 51 has the timing advanced by Δt, the timing adjusting unit 110 has the first timing as shown in FIG. By delaying the time axis (reference time) of the measurement data D1 output from the sensor 41 by Δt, the timings of the pulse portions included in each of the measurement data D1 to D3 are matched, and thereby between the measurement data D1 to D3. Adjust the timing of.

The end point detection unit 100 according to the present embodiment may be composed of one computer or a quantum computing system, or a plurality of computers or quantum computing systems connected to each other via a network, but may be composed of one or a plurality of computers. A program for realizing the end point detection unit 100 in a computer or a quantum computing system and a computer-readable recording medium in which the program is recorded non-transitory are also subject to the protection of the present case.

<Example of end point detection method>
Next, an example of the end point detection method by the end point detection unit 100 having such a configuration will be described. FIG. 18A is a flowchart showing an example of the end point detection method.

As shown in FIG. 18A, first, before the start of polishing, the timing adjusting unit 110 is applied to a plurality of types of end point detection sensors (first to third sensors 51 to 53) provided in one target polishing unit 3A. The timing synchronization signal is input at the same time (step S11).

Then, as shown in FIG. 16, the timing adjusting unit 110 is caused by the timing synchronization signal of the measurement data D1 to D3 output from each of the plurality of end point detection sensors (first to third sensors 51 to 53). The timings of the pulsed portions are compared (step S12). In the example shown in FIG. 16, the pulse portion of the measurement data D2 output from the second sensor 52 and the pulse portion of the measurement data D3 output from the third sensor 53 have the same timing, but are compared with them. , The pulse portion of the measurement data D1 output from the first sensor 51 is advanced (earlier) by Δt.

When the timings of the pulse portions of the measurement data D1 to D3 do not match (step S13: YES), the timing adjusting unit 110 adjusts the timing between the measurement data D1 to D3 so that the timings of the pulse portions match (step S13: YES). Step S14). For example, as shown in FIG. 16, when the pulse portion of the measurement data D1 output from the first sensor 51 has the timing advanced by Δt, the timing adjusting unit 110 has the first timing as shown in FIG. By delaying the time axis (reference time) of the measurement data D1 output from the sensor 41 by Δt, the timings of the pulse portions included in each of the measurement data D1 to D3 are matched, and thereby between the measurement data D1 to D3. Adjust the timing of.

Next, with reference to FIG. 10A, the polishing start output from each of the plurality of types of end point detection sensors (first to third sensors 51 to 53) provided in one target polishing unit 3A at the time of new polishing is started. The measurement data from to the present time is input to the trained model 121 of the determination unit 120 (step S15).

As a modification, referring to FIG. 11A, polishing output from each of a plurality of types of end point detection sensors (first to third sensors 51 to 53) provided in one target polishing unit 3A at the time of new polishing. Of the measurement data from the start to the present time, the temperature of the polishing pad 10 acquired from the start of polishing to the present time, the temperature of the slurry, the flow rate of the slurry, the pressure in each pressure chamber of the top ring 31A, and the number of times the polishing pad 10 has been used. One or more auxiliary information of may be input to the trained model 121 of the determination unit 120.

The determination unit 120 refers to FIG. 10A and starts from the start of polishing output from each of the plurality of types of end point detection sensors (first to third sensors 51 to 53) provided in one polishing unit 3A at the time of new polishing. Using the trained model 121 as input, it estimates and outputs whether or not the current polishing conditions are normal and whether or not the current time is the timing of the end point indicating the end of polishing. (Step S16).

For example, as shown in FIG. 12, the measurement data D1 to D3 from the start of polishing to the present time output from each of the first to third sensors 51 to 53 at the time of new polishing are from the start of polishing to the end of polishing, respectively. When the area A1 to A3 that should be statistically included when the polishing conditions are normal and the current time does not match the timing of the end of polishing, the determination unit 120 determines the current polishing. It is presumed that the conditions are normal and the current time is not the timing of the end point indicating the end of polishing, and the output is performed.

Further, for example, as shown in FIG. 13, the measurement data D1 to D3 from the start of polishing to the present time output from each of the first to third sensors 51 to 53 at the time of new polishing are obtained from the start of polishing to the end of polishing, respectively. When the polishing conditions up to are within the regions A1 to A3 that should be statistically included when the polishing conditions are normal, and the current time coincides with the timing of the end of polishing, the determination unit 120 determines the current time. It is estimated that the polishing conditions of the above are normal and the current time is the timing of the end point indicating the end of polishing, and the output is performed.

Further, for example, as shown in FIG. 14, any one of the measurement data D1 to D3 from the start of polishing to the present time output from each of the first to third sensors 51 to 53 at the time of new polishing (in the illustrated example). If the measurement data D1) of the first sensor 51 deviates from the region A1 that should be statistically included when the polishing conditions from the start to the end of polishing are normal, the determination unit 120 is at the present time. It is estimated that the polishing conditions of the above are abnormal and output.

As a modification, as shown in FIG. 11, the determination unit 120 outputs from each of the plurality of types of end point detection sensors (first to third sensors 51 to 53) provided in one polishing unit 3A at the time of new polishing. Whether the current polishing conditions are normal using the learned model 121 related to one modification by inputting the measured data from the start of polishing to the present time and the auxiliary information acquired from the start of polishing to the present time. You may estimate and output whether or not and whether or not the current time is the timing of the end point.

With reference to FIG. 14, when the determination unit 120 determines that the current polishing condition is normal (step S17: YES), and when it is determined that the current time is the timing of the end point (step S18). : YES), the first polishing stop unit 130 transmits a control signal for stopping the operation of the polishing unit 3A to the polishing unit 3A (step S19).

Further, referring to FIG. 13, when the determination unit 120 determines that the current polishing condition is normal (step S17: YES), and when it is determined that the current time is not the timing of the end point (step S17: YES). Step S18: NO), the end point detection unit 100 repeats the process from step S15.

On the other hand, referring to FIG. 15, when the determination unit 120 determines that the current polishing condition is abnormal (step S17: NO), the second polishing stop unit 140 stops the operation of the polishing unit 3A. A control signal is transmitted to the polishing unit 3A, and an alarm is issued (step S20).

By the way, as described above, conventionally, when a plurality of types of end point detection sensors are used in combination, which end point detection sensor measurement data is preferentially used by individually tuning each polishing unit. It was necessary for the operator to explicitly instruct when to switch the priority order between the end point detection sensors, and there was a lot of work for individual response, which required time and cost for individual response. In addition, there is insufficient accuracy for detecting the end point of a high-precision requirement corresponding to a fine pattern.

On the other hand, according to the above-described embodiment, each of the plurality of types of end point detection sensors (for example, the first to third sensors 51 to 53) provided in one polishing unit 3A outputs data at the time of new polishing. For the measured data, past polishing without the operator explicitly instructing which end point detection sensor's measurement data should be used with priority and when to switch the priority between end point detection sensors. By using the trained model 121 that machine-learned the waveform of the measurement data from the start of polishing to the end of polishing that was output at times, the current polishing conditions are considered in view of the similarity with the waveform of the measurement data at the time of past polishing. Can be estimated and output whether or not is normal and whether or not the current time is the timing of the end point. Therefore, it is possible to optimally combine and use the measurement data of a plurality of types of end point detection sensors (first to third sensors 51 to 53), and it is possible to improve the accuracy of end point detection.

Further, according to a modification of the present embodiment, measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors (first to third sensors 51 to 53) during past polishing. Wave and the temperature of the polishing pad 10 from the start of polishing to the end of polishing acquired during the past polishing, the temperature of the slurry, the flow rate of the slurry, the pressure in each pressure chamber of the top ring 31A, the number of times the polishing pad 10 has been used, etc. By using the trained model 121 that machine-learned the relationship with the auxiliary information of, the current polishing conditions are normal in view of the similarity between the waveform of the measurement data at the time of past polishing and the relationship of the auxiliary information. It is possible to estimate and output whether or not it is, and whether or not the current time is the timing of the end point. Therefore, it is possible to optimally combine and use the measurement data (first to third sensors 51 to 53) of the plurality of types of end point detection sensors and the auxiliary information, and it is possible to further improve the accuracy of end point detection.

Further, according to the present embodiment, the timing adjusting unit 110 adjusts the timing between the measurement data output from each of the plurality of types of end point detection sensors (first to third sensors 51 to 53), and then determines the determination unit. Since the input is input to the 120, the determination unit 120 can more accurately grasp the waveform of the measurement data output from each of the plurality of types of end point detection sensors (first to third sensors 51 to 53). This makes it possible to more accurately determine whether or not the current polishing conditions are normal and whether or not the current time is the timing of the end point. Therefore, the accuracy of end point detection can be further improved.

<Another example of end point detection method>
Next, another example of the end point detection method by the end point detection unit 100 will be described. FIG. 18B is a flowchart showing another example of the end point detection method. Of the end point detection methods shown in FIG. 18B, steps S11 to S14 are the same as the end point detection method shown in FIG. 18A, and description thereof will be omitted.

In the example shown in FIG. 18B, after step S14, with reference to FIG. 10B, from each of the plurality of types of end point detection sensors (first to third sensors 51 to 53) provided in one target polishing unit 3A. The measurement data from the start of polishing to the present time, which is output at the time of new polishing, is input to the trained model 121 of the determination unit 120 (step S151).

As a modification, referring to FIG. 11B, polishing output from each of a plurality of types of end point detection sensors (first to third sensors 51 to 53) provided in one target polishing unit 3A at the time of new polishing. Of the measurement data from the start to the present time, the temperature of the polishing pad 10 acquired from the start of polishing to the present time, the temperature of the slurry, the flow rate of the slurry, the pressure in each pressure chamber of the top ring 31A, and the number of times the polishing pad 10 has been used. One or more auxiliary information of may be input to the trained model 121 of the determination unit 120.

With reference to FIG. 10B, the determination unit 120 starts from the start of polishing output from each of the plurality of types of end point detection sensors (first to third sensors 51 to 53) provided in one polishing unit 3A at the time of new polishing. Using the trained model 121 as input, the measurement data up to the present time is used to estimate and output whether or not the polishing conditions at the present time are normal, and the remaining time from the present time to the timing of the end point indicating the end of polishing ( Step S161).

For example, as shown in FIG. 12, the measurement data D1 to D3 from the start of polishing to the present time output from each of the first to third sensors 51 to 53 at the time of new polishing are from the start of polishing to the end of polishing, respectively. When the area A1 to A3 that should be statistically included when the polishing conditions are normal and the current time does not match the timing of the end of polishing, the determination unit 120 determines the current polishing. It is estimated that the conditions are normal, and the remaining time Te from the present time to the timing of the end point indicating the end of polishing is estimated and output.

Further, for example, as shown in FIG. 13, the measurement data D1 to D3 from the start of polishing to the present time output from each of the first to third sensors 51 to 53 at the time of new polishing are obtained from the start of polishing to the end of polishing, respectively. When the polishing conditions up to are in the regions A1 to A3 that should be statistically included when the polishing conditions are normal, and the current time coincides with the timing of the end of polishing, the determination unit 120 determines the current time. It is estimated that the polishing conditions of the above are normal, and the remaining time Te from the present time to the timing of the end point indicating the end of polishing is estimated to be zero and output.

As a modification, as shown in FIG. 11, the determination unit 120 outputs from each of the plurality of types of end point detection sensors (first to third sensors 51 to 53) provided in one polishing unit 3A at the time of new polishing. Whether the current polishing conditions are normal using the learned model 121 related to one modification by inputting the measured data from the start of polishing to the present time and the auxiliary information acquired from the start of polishing to the present time. In addition to estimating whether or not, the remaining time Te from the current time to the timing of the end point may be estimated and output.

With reference to FIG. 14, when the determination unit 120 determines that the current polishing condition is normal (step S17: YES), and the remaining time Te estimated by the determination unit 120 has elapsed. (Step S181: YES), the first polishing stop unit 130 transmits a control signal for stopping the operation of the polishing unit 3A to the polishing unit 3A (step S19).

Further, referring to FIG. 13, the case where the determination unit 120 determines that the current polishing condition is normal (step S17: YES), and the remaining time Te estimated by the determination unit 120 has not elapsed. In the case (step S181: NO), the end point detection unit 100 repeats the process from step S15.

According to the above-described embodiment, measurement data output from each of a plurality of types of end point detection sensors (for example, first to third sensors 51 to 53) provided in one polishing unit 3A at the time of new polishing. Is output during past polishing without the operator explicitly instructing which end point detection sensor measurement data should be used with priority and when to switch the priority order between end point detection sensors. By using the trained model 121 in which the waveform of the measurement data from the start of polishing to the end of polishing is machine-learned, the current polishing conditions are normal in view of the similarity with the waveform of the measurement data at the time of past polishing. It is possible to estimate whether or not there is, and to estimate and output the remaining time Te from the present time to the timing of the end point. Therefore, it is possible to optimally combine and use the measurement data of a plurality of types of end point detection sensors (first to third sensors 51 to 53), and it is possible to improve the accuracy of end point detection.

Further, according to a modification of the embodiment, the measurement data from the start to the end of polishing output from each of the plurality of types of end point detection sensors (first to third sensors 51 to 53) during the past polishing. Corrugations, the temperature of the polishing pad 10 from the start of polishing to the end of polishing acquired during the past polishing, the temperature of the slurry, the flow rate of the slurry, the pressure in each pressure chamber of the top ring 31A, the number of times the polishing pad 10 has been used, etc. By using the trained model 121 in which the relationship with the auxiliary information is machine-learned, the current polishing conditions are normal in view of the similarity between the waveform of the measurement data at the time of past polishing and the relationship between the auxiliary information. It is possible to estimate whether or not there is, and to estimate and output the remaining time Te from the present time to the timing of the end point. Therefore, it is possible to optimally combine and use the measurement data (first to third sensors 51 to 53) of the plurality of types of end point detection sensors and the auxiliary information, and it is possible to further improve the accuracy of end point detection.

In the above-described embodiment, a plurality of types of end point detection sensors (first to first) provided in one polishing unit 3A are used as teacher data when the trained model 121 (tuned neural network system) is generated. The measurement data at each time point from the start of polishing to the end of polishing (that is, the data set of normal end) output from each of the third sensors 51 to 53) during the past "normal end" polishing was used, but the teacher data. Is not limited to the data set of normal termination, and may be a data set of abnormal termination, or may be a data set in which a data set of normal termination and a data set of abnormal termination are mixed. In training using a normally completed dataset, a trained model 121 (tuned neural network system) that determines a normal state or condition can be generated. In training using the abend data set, a trained model 121 (tuned neural network system) for determining an abnormal state or condition can be generated. In training using a data set in which a normal termination data set and an abnormal termination data set are mixed, a trained model 121 (tuned neural network system) for judging a normal state or condition and an abnormal state or condition is generated. it can. In this case, it is optimal to use the normal termination data set / abnormal termination data set = 70% / 30% to 99% / 1% as the ratio of the normal termination data set to the abnormal termination data set.

<Control of the entire board processing device>
Next, with reference to FIG. 19, the control of the entire substrate processing apparatus by the control unit 65 will be described. The control unit 65, which is the main controller, has a CPU, a memory, a recording medium, software recorded on the recording medium, and the like. The control unit 65 monitors and controls the entire board processing device, and performs signal transmission / reception, information recording, and calculation for that purpose. The control unit 65 mainly sends and receives signals to and from the unit controller 760. The unit controller 760 also has a CPU, a memory, a recording medium, software recorded on the recording medium, and the like. In the case of FIG. 19, the control unit 65 has a built-in program that functions as an end point detecting means for detecting the polishing end point indicating the end of polishing and a control means for controlling polishing by the polishing unit. The unit controller 760 may include a part or all of this program. The program can be updated. The program does not have to be updatable.

According to the embodiments described with reference to FIGS. 19 to 23, the following problems can be solved. The following points are problems of the control method of a typical polishing apparatus so far. Regarding the end point detection, a plurality of tests are performed before polishing the object, the polishing conditions and the end point determination conditions are obtained from the obtained data, and a recipe which is the polishing conditions is created. Although some signal analysis may be used, the wafer structure is processed to determine the end point detection by using one sensor signal. This did not provide sufficient accuracy for the following requirements. In order to improve the yield of the device or chip to be manufactured, it is necessary to detect the end point with higher accuracy in the manufacture of the device or chip and to suppress the variation between lots or chips to be small. In order to realize this, by using a system for detecting the end point to which the examples shown in FIGS. 19 to 23 are applied, it is possible to detect the end point with higher accuracy, and the yield is improved and the amount of polishing between chips varies. Can be reduced.

In particular, high-speed data processing, signal processing of many types and many sensors, standardized data of these signals, and data sets used for learning using artificial intelligence (AI) from data and determining end point detection. Realization of learning by creating and accumulating judgment examples by the created data set, improving accuracy by learning effect, polishing parameters judged and updated by the learned judgment function, and reflecting these polishing parameters in the high-speed control system. High-speed communication processing system, etc. can be realized. These are applicable to all the embodiments shown prior to FIG.

The unit controller 760 controls the units 762 (one or more) mounted on the board processing device. The unit controller 760 is provided for each unit 762 in the present embodiment. The unit 762 includes an unloading unit 62, a polishing unit 63, a cleaning unit 64, and the like. The unit controller 760 controls the operation of the unit 762, sends and receives signals to and from the monitoring sensor, sends and receives control signals, and performs high-speed signal processing and the like. The unit controller 760 is composed of an FPGA (field-programmable gate array), an ASIC (application specific integrated circuit, an integrated circuit for a specific application), and the like.

The unit 762 operates by a signal from the unit controller 760. Further, the unit 762 receives the sensor signal from the sensor and transmits it to the unit controller 760. The sensor signal may be further sent from the unit controller 760 to the control unit 65. The sensor signal is processed (including arithmetic processing) by the control unit 65 or the unit controller 760, and a signal for the next operation is sent from the unit controller 760. The unit 762 operates accordingly. For example, the unit controller 760 detects the torque fluctuation of the swing arm 110 by the current change of the swing shaft motor 14. The unit controller 760 sends the detection result to the control unit 65. The control unit 65 detects the end point.

The software includes, for example, the following. The software determines the type of polishing pad 10 and the amount of slurry supplied from the data recorded in the control device (control unit 65 or unit controller 760). Next, the software identifies the polishing pad 10 that can be used until the maintenance period or the maintenance period of the polishing pad 10, calculates the slurry supply amount, and outputs these. The software may be software that can be installed on the board processing device 764 after the board processing device 764 is shipped.

Communication between the control unit 65, the unit controller 760, and the unit 762 can be either wired or wireless. Communication via the Internet and other communication means (high-speed communication by a dedicated line) can be used with the outside of the board processing apparatus 764. Regarding data communication, it is possible to use the cloud by linking with the cloud, and to exchange data via the smartphone in the board processing device by linking with the smartphone. As a result, it is possible to exchange the operating status of the substrate processing apparatus and the setting information of the substrate processing with the outside of the substrate processing apparatus. As a communication device, a communication network may be formed between sensors and this communication network may be used.

It is also possible to perform automated operation of the board processing device by using the above control function and communication function. For automated operation, it is possible to standardize the control pattern of the substrate processing device and use the threshold value in determining the polishing end point.

It is possible to predict / judge / display the abnormality / life of the board processing device. It is also possible to perform control for performance stabilization.

Automatically extracts various data during operation of the substrate processing device and feature quantities of polishing data (thickness and end point of polishing) to automatically learn the operating state and polishing state, and automatically standardize control patterns. It is possible to predict / judge / display an abnormality / life.

It is possible to manage devices / devices by standardizing, for example, formats in communication methods, device interfaces, etc., and using them for information communication between devices / devices.

Next, in the board processing device 764, information is acquired from the semiconductor wafer 16 by a sensor, and a data processing device installed inside / outside the factory where the board processing device is installed via a communication means such as the Internet ( An embodiment in which data is stored in a cloud or the like, the data stored in the cloud or the like is analyzed, and the substrate processing apparatus is controlled according to the analysis result will be described. FIG. 20 shows the configuration of this embodiment.

1. 1. The information acquired from the semiconductor wafer 16 by the sensor can be as follows.
-Measurement signal or measurement data related to torque fluctuation of the oscillating shaft motor 14-Measurement signal or measurement data of SOPM (Spectrum Optical Endpoint Monitoring; optical sensor) -Measurement signal or measurement data of vortex current sensor-Top ring or polishing table Measurement signal or measurement data of vibration sensor that monitors vibration-Measurement signal or measurement data of sound sensor (not shown) that monitors changes in sound generated from the contact area between the wafer and the polishing pad-One or more of the above Combination measurement signal or measurement data

2. 2. The functions and configurations of communication means such as the Internet are as follows.
-The signal or data including the above-mentioned measurement signal or measurement data is transmitted to the data processing device 768 connected to the network 766.
-The network 766 may be a communication means such as the Internet or high-speed communication. For example, a network 766 connected in the order of a substrate processing device, a gateway, the Internet, a cloud, the Internet, and a data processing device is possible. High-speed communication includes high-speed optical communication, high-speed wireless communication, and the like. Further, as high-speed wireless communication, Wi-Fi (registered trademark), Bluetooth (registered trademark), Wi-Max (registered trademark), 3G, 4G, LTE, 5G and the like can be considered. Other high-speed wireless communication is also applicable. It is also possible to use the cloud as a data processing device.
-When the data processing device 768 is installed in the factory, it is possible to process signals from one or more board processing devices in the factory.
-When the data processing device 768 is installed outside the factory, it is possible to transmit signals from one or more board processing devices in the factory to the outside of the factory and process them. At this time, it is possible to connect to a data processing device installed in Japan or abroad.

3. 3. Regarding the data processing device 768 analyzing the data stored in the cloud or the like and controlling the board processing device 764 according to the analysis result, the following can be performed.
-After the measurement signal or measurement data is processed, it can be transmitted to the substrate processing apparatus 764 as a control signal or control data.
-The substrate processing device 764 that has received the data updates the polishing parameters related to the polishing process to perform the polishing operation based on the data, and the data from the data processing device 768 indicates that the end point has been detected. / In the case of data, it is judged that the end point has been detected, and polishing is finished. The polishing parameters include (1) pressing force on the four regions of the semiconductor wafer 16, that is, the central portion, the inner intermediate portion, the outer intermediate portion, and the peripheral portion, (2) the polishing time, and (3) the polishing table 30A and the top. There are the number of rotations of the ring 31A, (4) a threshold value for determining the polishing end point, and the like.

Next, another embodiment will be described with reference to FIG. FIG. 21 is a diagram showing a modified example of the embodiment of FIG. In this embodiment, the substrate processing device, the intermediate processing device, the network 766, and the data processing device are connected in this order. The intermediate processing device is composed of, for example, FPGA or ASIC, and has a filtering function, a calculation function, a data processing function, a data set creation function, and the like.

It is divided into the following three cases depending on how to use the Internet and high-speed optical communication. When (1) the Internet is between the board processing device and the intermediate processing device and the network 766 is the Internet, (2) high-speed optical communication is performed between the board processing device and the intermediate processing device, and the network 766 is high-speed. In the case of optical communication, (3) high-speed optical communication may be performed between the substrate processing device and the intermediate processing device, and the Internet may be outside the intermediate processing device.

Case (1): When the data communication speed and data processing speed in the entire system are sufficient as the Internet communication speed. The data sampling speed is about 1 to 1000 mS, and data communication of a plurality of polishing condition parameters can be performed. In this case, the intermediate processing device 770 creates a data set to be sent to the data processing device 768. The details of the data set will be described later. Upon receiving the data set, the data processing device 768 performs data processing, for example, calculates the change value of the polishing condition parameter up to the end point position, creates a process plan of the polishing process, and returns it to the intermediate processing device 770 through the network 766. The intermediate processing device 770 sends a change value of the polishing condition parameter and a necessary control signal to the substrate processing device 764.

In the case of (2): Communication between the board processing device and the intermediate processing device, and between the sensor signal between the intermediate processing device and the data processing device and between the state management devices is high-speed communication. In high-speed communication, communication is possible at a communication speed of 1 to 1000 Gbps. In high-speed communication, data, datasets, commands, control signals, etc. can be communicated. In this case, the intermediate processing device 770 creates a data set and transmits it to the data processing device 768. The intermediate processing device 770 extracts data necessary for processing in the data processing device 768, processes the data, and creates a data set. For example, a plurality of sensor signals for detecting the end point are extracted and created as a data set.

The intermediate processing device 770 sends the created data set to the data processing device 768 by high-speed communication. The data processing device 768 calculates the parameter change value up to the polishing end point and creates the process plan based on the data set. The data processing device 768 receives data sets from a plurality of board processing devices 764, calculates parameter update values for the next step and creates a process plan for each device, and uses the updated data sets as an intermediate processing device. Send to 770. Based on the updated data set, the intermediate processing device 770 converts the updated data set into a control signal and transmits the updated data set to the control unit 65 of the board processing device 764 by high-speed communication. The substrate processing apparatus 764 performs polishing in response to the updated control signal, and performs accurate end point detection.

In the case of (3): The intermediate processing device 770 receives a plurality of sensor signals of the board processing device 764 by high-speed communication. In high-speed optical communication, communication at a communication speed of 1 to 10000 Gbps is possible. In this case, it is possible to control online polishing conditions by high-speed communication between the substrate processing device 764, the sensor, the control unit 65, and the intermediate processing device 770. The data processing order is, for example, sensor signal reception (board processing device 764 to intermediate processing device 766), data set creation, data processing, parameter update value calculation, update parameter signal transmission, polishing control by the control unit 65, and update. The order is end point detection.

At this time, the intermediate processing device 770 performs high-speed end point detection control by the intermediate processing device 770 for high-speed communication. A status signal is periodically transmitted from the intermediate processing device 770 to the data processing device 768, and the control state monitoring process is performed by the data processing device 768. The data processing device 768 receives status signals from the plurality of board processing devices 764, and plans the next process process for each board processing device 764. A planning signal of the process process based on the plan is sent to each substrate processing apparatus 764, and each substrate processing apparatus 764 performs the preparation of the polishing process and the execution of the polishing process independently of each other. In this way, high-speed end point detection control is performed by the intermediate processing device 770 for high-speed communication, and state management of the plurality of board processing devices 764 is performed by the data processing device 768.

Next, an example of a data set will be described. It is possible to make a data set of sensor signals and required control parameters. The data set includes pressing the top ring 31A against the semiconductor wafer 16, the current of the swing shaft motor 14, the motor current of the polishing table 30A, the measurement signal of the optical sensor, the measurement signal of the eddy current sensor, and the polishing pad 10. The position / slurry of the top ring 31A, the flow / type of the chemical solution, the correlation calculation data thereof, and the like can be included.

The above types of data sets can be transmitted using a transmission system that transmits one-dimensional data in parallel or a transmission system that transmits one-dimensional data sequentially. As a data set, the above-mentioned one-dimensional data can be processed into two-dimensional data to form a data set. For example, assuming that the X-axis is time and the Y-axis is a large number of data strings, a plurality of parameter data at the same time are processed into one data set. Two-dimensional data can be treated as something like two-dimensional image data. This merit is that since the transfer of two-dimensional data is performed, the data can be exchanged and handled as time-related data with less wiring than the transfer of one-dimensional data. Specifically, if one-dimensional data is directly converted into one signal and one line, a large number of wires are required, but in the case of two-dimensional data transfer, a plurality of signals can be sent by one line. Further, when a plurality of lines are used, the interface with the data processing device 768 that receives the transmitted data becomes complicated, and the data reassembly in the data processing device 768 becomes complicated.

In addition, if there is a two-dimensional data set associated with such time, a comparison between the data set at the time of polishing under the standard polishing conditions performed before and the data set at the current standard polishing conditions is performed. Becomes easier. In addition, the differences between the two-dimensional data can be easily known by difference processing or the like. It is also easy to extract the difference and detect the sensor or parameter signal in which an abnormality occurs. In addition, it is easy to detect anomalies by extracting the parameter signal of the part where the difference from the surroundings is different by comparing the previous standard polishing conditions with the data set during the current polishing.

Next, another embodiment will be described with reference to FIG. FIG. 22 is a diagram showing a modified example of the embodiment of FIG. This embodiment is an example of a semiconductor factory. There are a plurality of substrate processing devices 764 in the factory. The substrate processing apparatus 764 that performs polishing and end point detection may have the same equipment and functions as those shown in FIGS. 19 to 21. For example, in end point detection using a large number of sensors (10 or more, the number of types ≥ 3), the amount of sensor signal data is large. At this time, if communication is performed using the Internet in order to create a data set and perform data analysis and update of polishing condition parameters, communication takes time. Therefore, the communication line L1 connecting the substrate processing device 764 and the intermediate processing device 770 is performed by using a high-speed communication device that performs high-speed optical communication, high-speed wireless communication, or the like. The intermediate processing device 770 is located near the sensor or the substrate processing device 764 and processes the signal from the sensor or the controller of the sensor at high speed. A signal for performing feedback or feedforward parameter update reflecting the processing result is transmitted to the substrate processing apparatus 764 at high speed. The substrate processing apparatus 764 receives the signal for updating the parameters, performs polishing processing, and detects the end point.

As shown in FIG. 22, when there are a plurality of substrate processing devices 764, there may be a first processing device 772 that receives signals from each substrate processing device 764 and performs processing in the factory. The first processing device 772 has a medium-sized memory and a calculation function, and can perform high-speed calculation. The first processing device 772 has an automatic learning function, performs automatic learning while accumulating data, and updates parameters for improving the uniformity of processing amount, improving the end point detection accuracy, and the like. By automatic learning, it is possible to continuously update the parameters to bring the parameters closer to the optimum values. In this case, high-speed communication is required when performing online processing in In situ, and the communication line L1 / communication line L2 is, for example, a high-speed optical communication communication line. At this time, for example, the data set creation is performed by the intermediate processing device 770, and the data analysis and parameter update can be performed by the first processing device 772. Then, a signal for reflecting the update parameter value to each board processing device 764 is sent to the board processing device 764 by the communication line L1 / communication line L2.

In addition, when high speed is not required, such as Inline monitoring, which measures uniformity while moving between polishing parts, the communication line L2 is relatively low speed, such as a communication line for Internet communication. In some cases, a communication line is sufficient. The initial polishing data is processed in the intermediate processing device 770, and the generated data set is sent to the first processing device 772 via the Internet. The first processing device 772 obtains analysis and parameter update values, and creates an update data set. The first processing device 772 sends it to the intermediate processing device 770. When the main polishing is performed in the next polishing unit, the update parameter value reflected from the update data set in the intermediate processing device 770 is sent to the substrate processing device 764, and the polishing is performed accordingly.

When exchanging information with the outside of the factory, the network 766 is used from the first processing device 772 to exchange data related to the information with the second processing device 774 outside the factory or a management device such as a personal computer. In this case, when communicating with the second processing device 774 outside the factory, the data related to the information may be encrypted in order to ensure security. Further, as the data related to the information, there is data indicating information related to the status of the substrate processing apparatus 764. In addition, data on the state of consumables of the board processing device 764 is exchanged, the replacement time is calculated by the second external processing device 774, and the customer is notified of the replacement time, or the board processing device 764 It can be displayed above.

Next, another embodiment will be described with reference to FIG. FIG. 23 is a diagram showing a modified example of the embodiment of FIG. This embodiment is an example of a semiconductor factory. There are a plurality of substrate processing devices 764 in the factory. The substrate processing apparatus 764 that performs polishing and end point detection may have the same equipment and functions as those shown in FIGS. 19 to 21. Compared with the embodiment of FIG. 22, the present embodiment is different in that there is a communication line L3 connected from the substrate processing device 764 to the first processing device 772 without going through the intermediate processing device 770. The feature of this embodiment is that communication using high-speed communication line L1 and communication line L2 is used for data communication that is created from data from a large number of sensors and forms a data set that requires high-speed communication for the creation. Is to use. Other control parameter communication that does not require high-speed communication is performed by connecting the board processing device 764 to the first processing device 772 via the communication line L3. For example, since a parameter group that does not require high-speed control can be used for the transport system, the cleaning system, the drying system, etc., for these systems, the substrate processing device 764 is used as the first processing device 772 by the communication line L3. Connect and do. Parameter signals and sensor signals that require high-speed communication, high-speed analysis, and high-speed communication data sets are variably selected according to the operating status of the board processing device 764, and the signals and the like are selected using the communication line L1 and the communication line L2. It may be transmitted and received.

In the present embodiment, data from the board processing device 764 is sent to the first processing device 772 in the factory using the communication line L2 and the communication line L3, and data analysis, automatic learning, parameter update value creation, and the like are performed. .. Then, the first processing device 772 sends the control parameters of the device in the next step to each substrate processing device 764. According to the present embodiment, when there are a plurality of board processing devices 764 in the factory, the first processing device 772 receives data from the plurality of board processing devices 764, processes the data, and makes each board processing device. The processing result can be sent to 764 via the intermediate processing apparatus 770.

As another embodiment in which this embodiment is modified, a configuration without a communication line L2 is also possible. The data related to the status of the high-speed processing state processed by the intermediate processing device 770 without using the communication line L2, together with the data related to the other device status status, is the first processing device via the communication line L3. It is possible to send to 772. In this case, the wiring for the communication line related to the communication line L2 can be reduced. That is, only where high-speed data processing and high-speed control are required, the high-speed communication line and the high-speed intermediate processing device 770 perform data processing, automatic learning, and control parameter update, and send the processing result to the board processing device 764. The status signals related to high-speed data processing and high-speed control and other status signals are sent together on the communication line L3 to the first processing device 772, and the first processing device 772 performs data processing, automatic learning, and control parameter update. Then, it is possible to send a signal including the processing result to each substrate processing apparatus 764. In the form shown in FIG. 23 and another form in which the same is changed, one first processing device 772 can handle a plurality of substrate processing devices 764. In these forms, the communication outside the factory is the same as that in FIG. 22.

Next, an example of the data set and automatic learning in the data processing and control modes shown in FIGS. 19 to 23 described above, and the calculation related thereto will be described. First, an example of a data set will be described. Regarding the data set, it is necessary to create a data set according to the processing in order to update effective control parameters as the processing such as polishing progresses. For example, for end point detection, it is preferable to use a data set in which a sensor signal that effectively captures the characteristics of the semiconductor film is selected. Using the polishing recipe, the recipe (polishing condition) corresponding to the film structure formed on the wafer is selected. At that time, it is possible to classify the membrane according to the following characteristics regarding the membrane structure. (1) Thinning the oxide film or insulating film, (2) Thinning the metal film or conductive film, (3) Polishing to the interface with the lower layer (the interface between the conductive layer and the insulating layer, etc.), (4) Polish the film-forming part to the pattern boundary (polishing of unnecessary parts after film-forming of wiring materials and insulating materials, etc.). Corresponding to this classification, the data set is a type of sensor suitable for the film when it is created by importing data from all types of sensors and when it is detected of the polishing state of the film. You may want to create a dataset by selecting data from.

There are the following data sets when creating by importing data from all types of sensors. For example, with data such as torque data (motor current, etc.) in TCM (motor current fluctuation measurement), torque data of arm with top ring (swing motor current, etc.), optical sensor (SOPM, etc.) data, eddy current sensor data, etc. , Create a data set that is a set of data obtained by calculating those data (differential data, etc.) and correlation data (high data such as absolute value of differentiated data and difference data of low data).

Regarding the detection of the polishing state of the film, there are the following data sets when creating a data set by selecting data from a sensor of the type suitable for the film. (1) When the oxide film or the insulating film is thinned, the value of the calculated data becomes high for the optical sensor signal having high sensitivity to the change in film thickness. In this case, by evaluating a plurality of data, for example, by adding the polishing time, the target polishing amount can be achieved and the end point is detected. For example, if the measured value by TCM and the arm torque data with the top ring are stable, it is considered that polishing at the same polishing rate is achieved. By changing the film thickness based on the optical sensor data, it is possible to accurately detect the end point by time counting based on the time when the film thickness reaches a certain thickness.

(2) When the metal film or the conductive film is thinned, the conductive film or the metal film is thinned, so that the calculated data of the eddy current sensor and the optical sensor, which are highly sensitive to the change in the film thickness of the conductive film, are used. It is used as a criterion for determining that the film thickness has reached a certain thickness. Similar to (1), when the measured value by TCM and the arm torque data with top ring are stable, the one with the higher calculated data value at the film thickness close to the target value is mainly selected, and the other is selected as the slave. .. The end point is detected by time counting based on the time when the film thickness reaches a certain thickness, mainly based on the change in film thickness based on the data of the selected sensor. Based on the data of the sensor selected as the slave, it is confirmed that there is no deviation (confirmation that the target area is almost reached), and the detection accuracy is improved.

As a method of using the data of the sensor selected as the slave, a priority ratio coefficient (weighting coefficient) is provided for the target values of both the main sensor and the slave sensor, and the main sensor and the slave sensor are used. It is also possible to specify the influence ratio, set the target value, and detect the end point. In addition, at this time, the data is used as learning data each time the number of times is repeated, and the judgment function is updated by learning in the judgment function (change of priority ratio coefficient, etc.) to improve the end point detection with higher accuracy. It is possible to continue.

(3) When polishing (over-polishing) to the interface with the lower layer, changes occur in all of the torque data in the TCM, the torque data of the arm with the top ring, the data of the optical sensor, and the data of the eddy current sensor. At this time, the torque data in the TCM and the arm torque data with the top ring generate a sudden change (pulse-like change) in the vicinity of the boundary surface in terms of the calculated data. Therefore, it is determined that the polishing region near the boundary surface is approached based on the data of the optical sensor and / or the data of the eddy current sensor. Next, it is possible to reset the end point detection time after a predetermined time has elapsed, based on the time when the change in the torque data and / or the torque data of the arm with the top ring is confirmed in the TCM. The reason for performing overpolishing in this way is as follows. When polishing to the boundary surface, if there is polishing residue, for example, if there is a metal-embedded vertical wiring, for example, if an oxide film remains on the bottom of the via or plug, the resistance value of the vertical wiring will increase and the circuit will malfunction. Causes. Therefore, overpolishing is performed so that there is no polishing residue. At the interface, the oxide film usually has small irregularities and is wavy before polishing. Therefore, considering the existence of small irregularities, it is necessary to perform overpolishing to remove the oxide film on the interface. Another reason for overpolishing is that the polishing device cannot be stopped abruptly when it reaches the interface. Therefore, over-polishing is performed with the elapse of the predetermined time as the end point detection time, and the polishing apparatus is stopped.

Resetting refers to the following processing method, for example. The threshold value of the signal waveform change amount of the arm torque data with the top ring is set as a provisional reference at the start of polishing, and a predetermined time is set as the number of counts of the remaining polishing time based on the time when the waveform is actually detected. It is possible to perform polishing by setting the count number as an update value of the end point detection time. At this time, among the torque data in the TCM and / or the torque data of the arm with the top ring, the one with the higher sensitivity is the main and the one with the lower sensitivity is the slave, and it is possible to process in the same manner as in (2). In order to improve the accuracy of resetting, learning can be used to set polishing parameters and update the set polishing parameters. In addition, a plurality of sensors can be used to improve the accuracy of resetting. As for learning, automatic learning is possible, but complex learning with some manuals is also possible.

(4) The same as (3) is used when the film-forming portion is polished to the pattern boundary portion (polishing of unnecessary parts after film formation of wiring material or insulating material, etc.). However, since the metal film and the insulating film are mixed in the film-forming portion, the fluctuation of the waveform after the boundary portion is larger than the others due to the influence of the pattern and the material of the boundary portion. It is difficult to detect the end point with only the eddy current sensor or the optical sensor. In such a case, it is effective to improve the accuracy by the data set created from the data of a plurality of sensors and the learning function using the data set, and to update the count number for end point detection by the priority ratio coefficient. When only one or two sensor signals are used, it is difficult to monitor the vicinity of the end point with high accuracy. Therefore, end point detection using multiple types (three or more types) of sensor data and a data set created from these data. Is very effective. When using such a large amount of data, learning improves the efficiency of accuracy improvement work.

In the case of (4), a data set is created using all the sensor signals related to end point detection, but it is also possible to select valid sensor data at the time of creating the data set and create the data set. This is particularly effective in the case of the simple film structures (1), (2), and (3).

Although the embodiments and modifications have been described above by way of example, the scope of the present technology is not limited to these, and can be changed or modified according to the purpose within the scope described in the claims. is there. In addition, each embodiment and modification can be appropriately combined as long as the processing contents do not contradict each other.

Claims

It has a learned model in which the waveforms of the measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors provided in one polishing unit during past polishing are machine-learned, and the plurality of types are described. It is equipped with a judgment unit that estimates and outputs whether or not the current time is the timing of the end point indicating the end of polishing by inputting the measurement data from the start of polishing to the present time output from each of the end point detection sensors at the time of new polishing. An end point detection device characterized by this.
It has a learned model in which the waveforms of the measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors provided in one polishing unit during past polishing are machine-learned, and the plurality of types are described. It is equipped with a judgment unit that estimates and outputs the remaining time from the current point to the timing of the end point indicating the end of polishing by inputting the measurement data from the start of polishing to the present time output from each of the end point detection sensors at the time of new polishing. An end point detection device characterized by.
The end point according to claim 1, further comprising a first polishing stop unit that transmits a control signal for stopping polishing to the polishing unit when the timing of the end point is estimated by the determination unit. Detection device.
The end point detecting device according to claim 3, further comprising a first polishing stop unit that transmits a control signal for stopping polishing to the polishing unit when the remaining time estimated by the determination unit has elapsed.
The determination unit inputs measurement data from the start of polishing to the present time output from each of the plurality of types of end point detection sensors at the time of new polishing, and estimates whether or not the present time is the timing of the end point indicating the end of polishing. The end point detection device according to claim 1 or 3, further comprising estimating and outputting whether or not the current polishing conditions are normal.
The first polishing stop unit transmits a control signal for stopping polishing to the polishing unit when it is estimated by the determination unit that the current polishing conditions are normal and the current time is the timing of the end point. The end point detecting device according to claim 5, wherein the characteristic claim 3 is cited.
The determination unit estimates the time from the present time to the timing of the end point indicating the end of polishing by inputting the measurement data from the start of polishing to the present time output from each of the plurality of types of end point detection sensors at the time of new polishing. The end point detecting device according to claim 2 or 4, wherein it estimates and outputs whether or not the current polishing conditions are normal.
The first polishing stop unit transmits a control signal for stopping polishing to the polishing unit when it is estimated by the determination unit that the current polishing conditions are normal and the remaining time is zero. The end point detecting device according to claim 7, wherein the characteristic claim 4 is cited.
The claim is further provided with a second polishing stop unit that transmits a control signal for stopping polishing to the polishing unit and issues an alarm when the determination unit determines that the current polishing state is abnormal. Item 4. The end point detecting device according to any one of Items 5 to 8.
The plurality of types of end point detection sensors are an optical sensor that shines light on an object to be polished and monitors a change in its reflectance, and a vortex current that applies a line of magnetic force to the object to be polished and monitors a change in the line of magnetic force due to a vortex current generated there. Vibration of the sensor, rocking torque sensor that monitors the change in torque applied to the rocking mechanism that swings the top ring, rotational torque sensor that monitors the change in torque applied to the rotating mechanism that rotates the polishing table, vibration of the top ring or polishing table The invention according to any one of claims 1 to 9, wherein the vibration sensor for monitoring and the sound sensor for monitoring the change in the sound generated from the contact portion between the object to be polished and the polishing pad are two or more types. End point detection device.
The trained model has a waveform of measurement data from each of the plurality of types of end point detection sensors output during past polishing from the start of polishing to the end of polishing, and from the start of polishing to the end of polishing acquired during the past polishing. Machine learning the relationship between the polishing pad temperature, slurry temperature, slurry flow rate, pressure in each pressure chamber of the top ring, and one or more auxiliary information of the number of times the polishing pad has been used.
The determination unit receives the measurement data from the start of polishing to the present time output from each of the plurality of types of end point detection sensors at the time of new polishing and the auxiliary information acquired from the start of polishing to the present time, and the present time is The end point detecting device according to claim 1, wherein the timing of the end point is estimated and output.
The trained model has a waveform of measurement data from each of the plurality of types of end point detection sensors output during past polishing from the start of polishing to the end of polishing, and from the start of polishing to the end of polishing acquired during the past polishing. Machine learning the relationship between the polishing pad temperature, slurry temperature, slurry flow rate, pressure in each pressure chamber of the top ring, and one or more auxiliary information of the number of times the polishing pad has been used.
The determination unit receives the measurement data from the start of polishing to the present time output from each of the plurality of types of end point detection sensors at the time of new polishing and the auxiliary information acquired from the start of polishing to the present time as input from the present time. The end point detection device according to claim 2, wherein the remaining time until the end point timing indicating the end of polishing is estimated and output.
The invention according to any one of claims 1 to 12, further comprising a timing adjusting unit that adjusts the timing between the measurement data output from each of the plurality of types of end point detection sensors and then inputs the timing to the determination unit. End point detection device.
The timing adjusting unit simultaneously inputs a timing synchronization signal to the plurality of types of end point detection sensors, and determines the timing of a pulse portion caused by the timing synchronization signal in the measurement data output from each of the plurality of end point detection sensors. The end point detection device according to claim 13, wherein the timings between the measurement data are matched by matching the measurement data.
The trained model machine-learns the waveform of the measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors provided in the first polishing unit during the past polishing. Machine learning the waveform of the measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors provided in the second polishing unit different from the first polishing unit during the past polishing. The end point detecting device according to any one of claims 1 to 14, wherein the end point detecting device is provided.
The end point detecting device according to claim 15, wherein the first polishing unit and the second polishing unit are installed in the same factory.
The end point detecting device according to claim 15, wherein the first polishing unit and the second polishing unit are installed in different factories.
Using a learned model in which the waveforms of the measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors provided in one polishing unit during past polishing are machine-learned, the plurality of types are described. Includes a judgment step that estimates and outputs whether or not the current time is the timing of the end point indicating the end of polishing by inputting the measurement data from the start of polishing to the present time output from each of the end point detection sensors at the time of new polishing. An end point detection method characterized by.
Using a learned model in which the waveforms of the measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors provided in one polishing unit during past polishing are machine-learned, the plurality of types are described. It includes a judgment step that estimates and outputs the remaining time from the current point to the timing of the end point indicating the end of polishing by inputting the measurement data from the start of polishing to the present time output from each of the end point detection sensors at the time of new polishing. Characteristic end point detection method.
Computer,
It has a learned model in which the waveforms of the measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors provided in one polishing unit during past polishing are machine-learned, and the plurality of types are described. The measurement data from the start of polishing to the present time, which is output from each of the end point detection sensors at the time of new polishing, is input, and it is made to function as a judgment unit that estimates and outputs whether or not the present time is the timing of the end point indicating the end of polishing. An end point detection program characterized by this.
Computer,
It has a learned model in which the waveforms of the measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors provided in one polishing unit during past polishing are machine-learned, and the plurality of types are described. The measurement data from the start of polishing to the present time, which is output from each of the end point detection sensors at the time of new polishing, is input, and the remaining time from the present time to the timing of the end point indicating the end of polishing is estimated and output as a judgment unit. An end point detection program featuring.
Computer,
It has a learned model in which the waveforms of the measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors provided in one polishing unit during past polishing are machine-learned, and the plurality of types are described. The end point that functions as a judgment unit that estimates and outputs whether or not the current time is the timing of the end point indicating the end of polishing by inputting real-time measurement data during polishing output from each of the end point detection sensors during new polishing. A computer-readable recording medium on which the detection program is recorded.
Computer,
It has a learned model in which the waveforms of the measurement data from the start of polishing to the end of polishing output from each of the plurality of types of end point detection sensors provided in one polishing unit during past polishing are machine-learned, and the plurality of types are described. The end point that functions as a judgment unit that estimates and outputs the remaining time from the current point to the timing of the end point indicating the end of polishing by inputting the measurement data from the start of polishing to the present time output from each of the end point detection sensors at the time of new polishing. A computer-readable recording medium on which the detection program is recorded.
It has an input layer, one or more intermediate layers connected to the input layer, and an output layer connected to the intermediate layer, and is past from each of a plurality of types of end point detection sensors provided in one polishing unit. The measurement data from the start of polishing to each time point output during polishing is input to the input layer, and the output result output from the output layer is compared with the information on whether or not the time point is the end point timing. By repeating the process of updating the parameters of each node according to the error for the measurement data from the start of polishing in the past polishing to each time point, each of the plurality of types of end point detection sensors outputs during the past polishing. This is a machine-learned waveform of the measurement data from the start of polishing to the end of polishing.
When the measurement data from the start of polishing to the present time, which is output from each of the plurality of types of end point detection sensors at the time of new polishing, is input to the input layer, it is estimated whether or not the present time is the timing of the end point indicating the end of polishing. A trained model for making a computer work so that it outputs from the output layer.
It has an input layer, one or more intermediate layers connected to the input layer, and an output layer connected to the intermediate layer, and is past from each of a plurality of types of end point detection sensors provided in one polishing unit. The measurement data from the start of polishing to each time point output during polishing is input to the input layer, and the output result output from the output layer and the information on the remaining time from that time point to the timing of the end point indicating the end of polishing By repeating the process of comparing the above and updating the parameters of each node according to the error for the measurement data from the start of polishing to each time point during the past polishing, each of the plurality of types of end point detection sensors in the past This is a machine-learned waveform of the measurement data output from the start of polishing to the end of polishing, which is output during polishing.
When the measurement data from the start of polishing to the present time, which is output from each of the plurality of types of end point detection sensors at the time of new polishing, is input to the input layer, the remaining time from the present time to the timing of the end point indicating the end of polishing is estimated. A trained model for making a computer work so that it outputs from the output layer.