JP2023122817A - Polishing end point detection method, polishing end point detection system, polishing device, and computer readable recording medium - Google Patents

Abstract

To provide a method for accurately predicting polishing end point time.SOLUTION: A method acquires a polishing monitoring value indicating polishing progress of a workpiece W at the time of polishing of the workpiece W, generates time-series data indicating a change of the polishing monitoring value along polishing time of the workpiece W, inputs polishing monitoring values VF1 and VF2 and polishing times TF1 and TF2 constituting feature data points FP1 and FP2 included in the time-series data to a learned model 34, and outputs a polishing end point prediction time from the learned model 34.SELECTED DRAWING: Figure 7

Description

The present invention relates to a method and system for detecting polishing endpoints of workpieces used in the manufacture of semiconductor devices such as wafers, substrates, and panels. The present invention also relates to a polishing apparatus equipped with such a polishing endpoint detection system. Furthermore, the present invention relates to a computer-readable recording medium recording a program for causing a polishing apparatus to execute a polishing endpoint detection method.

A polishing module typified by a CMP apparatus is used in the manufacture of semiconductor devices. A wiring structure of a semiconductor device is formed by forming a metal film (such as a copper film) on an insulating film in which wiring grooves are formed, and then removing unnecessary metal films using a polishing module. The polishing module polishes the surface of the workpiece by relatively moving the workpiece and the polishing pad while supplying a polishing liquid (slurry) to the polishing pad on the polishing table.

The polishing module includes a polishing endpoint detector that detects the polishing endpoint of the workpiece. This polishing end point detection device monitors polishing of the workpiece based on polishing index values indicating the progress of polishing (for example, table torque current, output signal of eddy current film thickness sensor, output signal of optical film thickness sensor). Then, the polishing end point, which is the point at which the film to be polished (for example, the metal film) is removed, is determined.

The polishing index value changes depending on the surface state of the workpiece as the polishing progresses. The polishing end point is detected, for example, based on the characteristic change point of the polishing index value due to the removal of the film to be polished and the exposure of the underlying layer. The polishing module terminates polishing of the workpiece based on the detected polishing endpoint.

JP 2018-58197 A JP 2013-176828 A

Since the measured values of the polishing index usually contain noise, the average of the measured values obtained over several seconds, that is, the moving average is used to detect the polishing end point. Due to the calculation of this moving average, the polishing end point may be detected after a delay from the time when the characteristic change point of the polishing index value actually appears.

Accordingly, the present invention provides a method and system that can accurately predict the polishing endpoint time.

In one aspect, when the workpiece is polished, a polishing monitoring value indicating the polishing progress of the workpiece is obtained, time-series data representing changes in the polishing monitoring value along the polishing time of the workpiece is generated, and the A polishing endpoint detection method is provided in which polishing monitoring values and polishing times that constitute feature data points included in time-series data are input to a trained model, and a polishing endpoint prediction time is output from the trained model.
In one aspect, the polishing monitoring value and the polishing time that constitute the initial polishing data point included in the time-series data are further input to the learned model, and the predicted polishing end point time is output from the learned model.
In one aspect, a polishing monitoring value and a polishing time constituting an intermediate data point between the initial polishing data point and the characteristic data point, which are included in the time-series data, are further input to the trained model, and Output the polishing end point prediction time from the trained model.

In one aspect, the feature data points are a plurality of feature data points included in the time-series data.
In one aspect, the polishing endpoint predicting device further inputs a polishing monitoring value and a polishing time that constitute intermediate data points between the plurality of feature data points included in the time-series data into the trained model. , output the polishing end point prediction time from the learned model.
In one aspect, at least one of the total usage time of the polishing pad used in the polishing module, the type of the polishing pad, the type of polishing liquid used for polishing the workpiece, and the type of the workpiece. Auxiliary information is further input to the learned model, and a polishing endpoint prediction time is output from the learned model.
In one aspect, the feature data point is any one of a maximum point, a minimum point, and an inflection point of the time-series data.
In one aspect, the characteristic data points are data points at which predetermined polishing monitoring values are reached.
In one aspect, the polishing progress meter is one of a torque current detector and a film thickness measurement sensor.

In one aspect, a polishing progress measuring device provided in a polishing module and a polishing endpoint predicting device having a learned model constructed by machine learning are provided, and the polishing endpoint predicting device measures the polishing progress during polishing of a workpiece. obtain a polishing monitoring value output from a device that indicates the polishing progress of the workpiece, generate time-series data representing changes in the polishing monitoring value along the polishing time of the workpiece, and store the time-series data in the A polishing endpoint detection system is provided configured to input polishing monitoring values and polishing times comprising feature data points included into said trained model and to output a predicted polishing endpoint time from said trained model. .
In one aspect, the polishing endpoint prediction device further inputs a polishing monitoring value and a polishing time constituting initial polishing data points included in the time-series data to the learned model, and predicts the polishing endpoint from the learned model. Configured to output time.
In one aspect, the polishing endpoint predicting device predicts polishing monitoring values and polishing times, which constitute intermediate data points between the initial polishing data points and the feature data points, included in the time-series data, and the learned model. and outputs the polishing end point prediction time from the learned model.

In one aspect, the feature data points are a plurality of feature data points included in the time-series data.

In one aspect, the polishing endpoint predicting device further inputs a polishing monitoring value and a polishing time that constitute intermediate data points between the plurality of feature data points included in the time-series data into the trained model. , to output a polishing end point prediction time from the learned model.
In one aspect, the polishing endpoint predicting device predicts the total usage time of the polishing pad used in the polishing module, the type of the polishing pad, the type of polishing liquid used for polishing the workpiece, and the type of polishing liquid used for polishing the workpiece. Auxiliary information of at least one piece type is further input to the learned model, and a polishing end point prediction time is output from the learned model.
In one aspect, the feature data point is any one of a maximum point, a minimum point, and an inflection point of the time-series data.
In one aspect, the characteristic data points are data points at which predetermined polishing monitoring values are reached.
In one aspect, the polishing progress meter is one of a torque current detector and a film thickness measurement sensor.

In one aspect, there is provided a polishing apparatus comprising a polishing table for supporting a polishing pad, a polishing head for polishing the workpiece by pressing the workpiece against the polishing surface of the polishing pad, and the polishing end point detection system. be done.
In one aspect, during polishing of a workpiece, obtaining a polishing monitoring value indicating polishing progress of the workpiece; and generating time-series data representing changes in the polishing monitoring value along polishing time of the workpiece. and inputting a polishing monitoring value and a polishing time that constitute characteristic data points included in the time-series data to a trained model, and outputting a predicted polishing end point time from the trained model. A recorded computer-readable recording medium is provided.

According to the present invention, by inputting a polishing monitoring value and a polishing time that constitute characteristic data points appearing before the polishing end point into the learned model, the predicted polishing end point time is output from the learned model. As a result, the polishing end point time can be accurately predicted before the target polishing end point, and polishing of the workpiece can be finished at the target polishing end point.

1 is a schematic diagram illustrating an embodiment of a polishing apparatus including a polishing endpoint detection system and a polishing module; FIG. FIG. 5 is a diagram showing an example of a graph showing time-series data representing changes in polishing monitoring values with polishing time of a workpiece; FIG. 5 is a diagram showing an example of a graph showing time-series data representing changes in polishing monitoring values including noise; FIG. 4 is an example graph showing time-series data generated during polishing of a learning workpiece; FIG. FIG. 4 is a schematic diagram showing an example of a trained model constructed using deep learning; FIG. 5 is a diagram illustrating data points at which predetermined polishing monitoring values are reached; 4 is an example graph showing time-series data generated during polishing of a workpiece; 4 is a flow chart describing one embodiment of a method for predicting the polishing endpoint of a workpiece using a trained model. It is a figure explaining the intermediate data point contained in time-series data. FIG. 4 is a schematic diagram showing another embodiment of a polishing endpoint detection system; FIG. 10 is a schematic diagram showing still another embodiment of the polishing endpoint detection system; FIG. 10 is a schematic diagram showing still another embodiment of the polishing endpoint detection system;

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described with reference to the drawings.
FIG. 1 is a schematic diagram showing an embodiment of a polishing apparatus having a polishing end point detection system 7 and a polishing module 1. As shown in FIG. The polishing module 1 is a polishing apparatus (CMP apparatus) that chemically and mechanically polishes the workpiece W. As shown in FIG. Examples of workpieces W include wafers, substrates, panels, etc. used in the manufacture of semiconductor devices. The polishing endpoint detection system 7 is configured to detect the polishing endpoint of the workpiece W being polished by the polishing module 1 .

A polishing module 1 includes a polishing table 3 for supporting a polishing pad 2 and a polishing head 6 connected to the lower end of a head shaft 5 . The head shaft 5 is connected to a polishing head rotating motor 12 via a torque transmission mechanism 10 such as a belt and pulleys. The polishing head rotating motor 12 rotates the head shaft 5 via the torque transmission mechanism 10, and the rotation of the head shaft 5 causes the polishing head 6 to rotate about the head shaft 5 in the direction indicated by the arrow. The polishing head 6 is configured to be able to hold a workpiece (eg wafer) W on its lower surface by vacuum suction.

The head shaft 5 is rotatably held by the head arm 15 . The head arm 15 is arranged above the polishing pad 2 and parallel to the polishing surface 2 a of the polishing pad 2 . One end of the head arm 15 holds the head shaft 5 and the other end of the head arm 15 is connected to an arm swing motor 16 . The arm swing motor 16 is configured to move the polishing head 6 between a polishing position on the polishing table 3 and a transfer position outside the polishing table 3 by rotating the head arm 15 by a predetermined angle. It is

The polishing table 3 is connected via a table shaft 3a to a table rotating motor 18 arranged below the table rotating motor 18. The table rotating motor 18 rotates the polishing table 3 about the table shaft 3a in the direction indicated by the arrow. It's like The upper surface of the polishing pad 2 constitutes a polishing surface 2a for polishing the workpiece W. As shown in FIG. A polishing liquid supply nozzle 20 for supplying polishing liquid (slurry) to the polishing surface 2a is arranged above the polishing table 3 .

Polishing of the workpiece W is performed as follows. While the polishing table 3 and the polishing pad 2 are rotated by the table rotation motor 18 , the polishing liquid is supplied from the polishing liquid supply nozzle 20 to the polishing surface 2 a of the polishing pad 2 . The polishing head 6 presses the workpiece W against the polishing surface 2 a of the polishing pad 2 while being rotated by the polishing head rotating motor 12 . The workpiece W is polished by a combination of the chemical action of the polishing liquid and the mechanical action of the abrasive grains contained in the polishing liquid and/or the polishing pad 2 . In one embodiment, the arm swing motor 16 may oscillate the polishing head 6 on the polishing surface 2a of the polishing pad 2 during polishing of the workpiece W. FIG.

Next, the polishing end point detection system 7 will be described. The polishing endpoint detection system 7 includes a polishing progress measuring device 30 arranged in the polishing module 1 and a polishing endpoint prediction device 33 for predicting the polishing endpoint of the workpiece W. FIG. In this embodiment, the polishing progress measuring device 30 is a torque current detector (that is, a table torque current detector) connected to the table rotation motor 18 . The polishing progress measuring device 30, which is a torque current detector, is configured to measure the current supplied to the table rotation motor 18 (hereinafter referred to as torque current). A torque current is a current required for the table rotation motor 18 to rotate the polishing table 3 and the polishing pad 2 at a constant speed.

The measured value of the torque current functions as a polishing monitoring value that indicates the polishing progress of the workpiece W. That is, since the surface of the workpiece W and the polishing surface 2a of the polishing pad 2 are in sliding contact with each other during polishing of the workpiece W, a frictional force is generated between the workpiece W and the polishing pad 2. FIG. This frictional force acts on the table rotation motor 18 as resistance torque. The frictional force generated between the workpiece W and the polishing pad 2 varies depending on the surface condition of the workpiece W. For example, when the surface of the work piece W has a wiring pattern structure, the frictional force changes depending on the removal state of the film or metal wiring forming the wiring pattern structure. This change in frictional force can be detected as a change in the torque current supplied to the table rotation motor 18 to rotate the polishing table 3 and polishing pad 2 at a preset speed. The polishing progress meter 30 is connected to the polishing endpoint predictor 33 and is configured to generate polishing monitoring values that are measurements of torque current and to send the polishing monitoring values to the polishing endpoint predictor 33 .

The polishing end point predicting device 33 includes a storage device 33a that stores a program and a learned model 34, which will be described later, and a computing device 33b that performs computation according to instructions included in the program. The polishing end point prediction device 33 is composed of at least one computer. The storage device 33a includes a main storage device such as a random access memory (RAM) and an auxiliary storage device such as a hard disk drive (HDD) and solid state drive (SSD). Examples of the arithmetic unit 33b include a CPU (central processing unit) and a GPU (graphic processing unit). However, the specific configuration of the polishing end point prediction device 33 is not limited to these examples.

The polishing endpoint prediction device 33 may be an edge server connected to the polishing progress measuring device 30 via a communication line, or a cloud server connected to the polishing progress measuring device 30 via a network such as the Internet. The polishing end point prediction device 33 may be a combination of multiple servers. For example, the polishing endpoint prediction device 33 may be a combination of an edge server and a cloud server connected to each other by a communication network such as the Internet or a local area network. However, the specific configuration of the polishing end point prediction device 33 is not limited to these examples. In one embodiment, the polishing end point predicting device 33 may be composed of a plurality of devices each having a storage device and an arithmetic device.

The polishing end point prediction device 33 is connected to the polishing control section 40 . The polishing end point prediction device 33 predicts the polishing end point of the workpiece W and sends a polishing end signal including the polishing end point prediction time to the polishing control section 40 . Based on the received polishing end signal, the polishing control unit 40 issues a command to the polishing module 1 to end the polishing of the workpiece W at the polishing end point prediction time.

FIG. 2 is a diagram showing an example of a graph showing time-series data representing changes in polishing monitoring values with polishing time of the workpiece W. As shown in FIG. In FIG. 2, the vertical axis represents the polishing monitoring value (ie, the torque current measurement), and the horizontal axis represents the polishing time of the workpiece W. As shown in FIG. As the workpiece W is polished, the surface condition of the workpiece W changes. The change in the surface condition of the work piece W appears as a change in the polishing monitoring value, as shown in FIG.

The polishing end point of the workpiece W is detected based on the polishing monitoring value representing the polishing progress of the workpiece W. FIG. The time-series data shown in FIG. 2 includes local maximum points MP of the polishing monitoring values. This maximum point MP is, for example, a characteristic data point that appears when the film to be polished is removed and the underlying layer is exposed. In the present embodiment, the target polishing end point is the time when the maximum point MP appears. Therefore, the polishing module 1 aims to finish polishing the workpiece W at the polishing end point time TE corresponding to the maximum point MP.

However, when the polishing end point is detected after the local maximum point MP appears, the polishing control unit 40 issues a command to the polishing module 1, and the time at which the polishing module 1 finishes polishing the workpiece W is past the polishing end point time TE. Sometimes I end up In addition, since the measured values of the polishing monitoring values contain noise, the waveform of the time-series data is unstable as shown in FIG. 3, making it impossible to accurately detect the polishing end point. Therefore, time-series data obtained by calculating a moving average of polishing monitoring values and reducing noise as shown in FIG. 2 is used to detect the polishing end point. Calculation of this moving average may delay the detection of the polishing end point.

Therefore, the polishing end point prediction device 33 of the present embodiment predicts the polishing end point using the learned model 34 constructed by machine learning based on the time-series data obtained when the learning workpiece is polished. do. As shown in FIG. 1, the polishing end point predicting device 33 has a learned model 34 stored in a storage device 33a.

The training workpiece used to build the trained model 34 may be of the same type as the workpiece W to be polished, or may be of a different type. The types of workpieces are distinguished by their structural characteristics, such as the surface structure of the workpiece, the material that constitutes the workpiece, and the size of the workpiece. The surface structure of the workpiece varies depending on the wiring pattern formed on the workpiece surface and the material forming the wiring pattern.

FIG. 4 is an example of a graph showing time-series data generated during polishing of a learning workpiece. The learning workpiece is polished by the polishing module 1 in the same manner as the polishing of the workpiece W described above. The learning workpiece is polished under the same polishing conditions as the workpiece W. Polishing conditions include the rotational speed of the polishing table 3, the rotational speed of the polishing head 6, the pressing force of the polishing head 6 against the workpiece W, and the flow rate of the polishing liquid.

A polishing progress meter 30 measures the torque current during polishing of the learning workpiece. The polishing end point predicting device 33 acquires the measured value of the torque current indicating the polishing progress of the learning workpiece, that is, the polishing monitoring value from the polishing progress measuring device 30, and calculates the polishing monitoring value along the polishing time of the learning workpiece. Generating time series data representing changes (ie, changes in torque current measurements). This time-series data is stored in the storage device 33 a of the polishing end point prediction device 33 .

In one embodiment, the polishing endpoint prediction device 33 calculates a first derivative, a second derivative, or a higher-order derivative of the measured value of the torque current to determine the polishing time of the learning workpiece. It is also possible to generate time-series data representing changes in polishing monitoring values, which are first derivative, second derivative, or higher-order derivative values of measured torque current values.

The time-series data shown in FIG. 4 includes an initial polishing data point IP, two feature data points FP1 and FP2, and a target endpoint data point EP. Each data point is a time-series data point composed of a polishing monitoring value and a corresponding polishing time.

The initial polishing data point IP is the data point when polishing of the learning workpiece is started. The initial polishing data point IP is composed of the polishing monitoring value VI and the polishing time TI. Since the polishing monitoring value at the start of polishing is unstable, the initial polishing data point IP may be the data point at the start of polishing (that is, when the polishing time TI = 0), or a predetermined value from the start of polishing. It may be a data point when the initial polishing time has elapsed.

The feature data points FP1 and FP2 are data points having predetermined features included in the time-series data, and appear before the target polishing end point (that is, the target end point data point EP). A data point with a predetermined characteristic is a data point that exhibits a characteristic change, such as a maximum point, minimum point, or inflection point, or a data point at which a predetermined polishing monitoring value is reached. Although there are two feature data points in the time series data shown in FIG. 4, there is one feature data point appearing before the polishing endpoint (i.e., the target endpoint data point EP) in the time series data for different types of workpieces. There may be only one, or three or more.

In the time-series data shown in FIG. 4, the feature data point FP1 is the maximum point and the feature data point FP2 is the minimum point. Characteristic data point FP1 is composed of polishing monitoring value VF1 and polishing time TF1. Characteristic data point FP2 is composed of polishing monitoring value VF2 and polishing time TF2. The target endpoint data point EP is a data point at the target polishing endpoint. The target endpoint data point EP is composed of a polishing monitoring value VE and a polishing time (polishing endpoint time) TE.

In the present embodiment, the training data used to build the learned model 34 includes polishing monitoring values and polishing times that constitute polishing initial data points IP and feature data points FP1 and FP2, and are correct labels. The polishing time (polishing end point time) TE that constitutes the end point data point EP is included. The trained model 34 is constructed by machine learning. Examples of machine learning include SVR method (support vector regression method), PLS method (partial least squares method: Partial Least Squares), deep learning method (deep learning method), random forest method, and decision tree method. . In one example, trained model 34 consists of a neural network constructed by deep learning methods.

FIG. 5 is a schematic diagram showing an example of a trained model 34 constructed using the deep learning method. The trained model 34 has an input layer 101 , multiple hidden layers (also called intermediate layers) 102 , and an output layer 103 . The trained model 34 shown in FIG. 5 has four hidden layers 102, but the configuration of the trained model 34 is not limited to the embodiment shown in FIG.

Construction of the trained model 34 using deep learning is performed as follows. The input layer 101 shown in FIG. 5 is input with polishing monitoring values and polishing times that form the initial polishing data point IP and the feature data points FP1 and FP2. More specifically, the polishing monitoring value VI and the polishing time TI that constitute the initial polishing data point IP, the polishing monitoring value VF1 and the polishing time TF1 that constitute the characteristic data point FP1, and the polishing monitoring that constitutes the characteristic data point FP2. A value VF2 and a polishing time TF2 are entered into the input layer 101 .

The trained model 34 is configured such that when the input layer 101 receives the polishing monitoring values and the polishing time that form the initial polishing data point IP and the characteristic data points FP1 and FP2, the output layer 103 outputs the predicted polishing end point time. is configured to The polishing end point prediction device 33 compares the polishing end point prediction time output from the output layer 103 with the polishing time (polishing end point time) TE that constitutes the target end point data point EP that is the correct label, and minimizes the error. Adjust the parameters (weights, thresholds, etc.) of each node (neuron) so that As a result, the learned model 34 is learned to output an appropriate polishing end point prediction time from the output layer 103 based on the data input to the input layer 101 .

A learned model 34 is constructed by repeating the above-described learning using a plurality of time-series data obtained when polishing a plurality of learning workpieces. The trained model 34 based on machine learning basically has low prediction accuracy for input data that has never been experienced. Therefore, the accuracy of the polishing endpoint prediction time output from the trained model 34 can be increased by using a large number of learning workpieces with polishing monitoring values that change in different manners.

In the embodiment described above, the initial polishing data point IP and the two feature data points FP1, FP2 are included in the training data. Only FP2 may be used. Training data may include all feature data points that occur up to the polishing endpoint. For example, if there are three feature data points appearing by polishing endpoint, the training data may include all three feature data points.

In other embodiments, the characteristic data points may be data points at which predetermined polishing monitoring values are reached. For example, as shown in FIG. 6, a characteristic data point FP is a data point when a predetermined polishing monitoring value VF is reached, and the characteristic data point FP is composed of a predetermined polishing monitoring value VF and a polishing time TF. ing. In this case, the training data used to build the learned model 34 includes polishing monitoring values and polishing times that constitute polishing initial data points IP and feature data points FP, and also target endpoint data points EP, which are correct labels. It includes the constituent polishing time (polishing end point time) TE.

In yet another embodiment, the trained model 34 may be constructed by inputting the time-series data from the start of polishing to the end of polishing into the input layer 101 of the trained model 34 . That is, the training data may include the time-series data itself from the start of polishing to the end point of polishing, and the correct label included in the training data may be the polishing time (polishing end point time) TE that constitutes the target end point data point EP. .

Next, a method of predicting the polishing end point of the workpiece W to be polished using the learned model 34 will be described. FIG. 7 is an example of a graph showing time-series data generated when the workpiece W is polished. The polishing progress meter 30 measures the torque current while the workpiece W is being polished. The polishing end point predicting device 33 acquires the measured value of the torque current indicating the polishing progress of the workpiece W, that is, the polishing monitoring value from the polishing progress measuring device 30, and changes the polishing monitoring value along the polishing time of the workpiece W ( That is, it generates time-series data representing changes in measured values of torque current). This time-series data is stored in the storage device 33 a of the polishing end point prediction device 33 .

In one embodiment, the polishing end point predicting device 33 calculates the first derivative value, the second derivative value, or the derivative value of a higher order of the measured value of the torque current, and follows the polishing time of the workpiece W Time-series data representing changes in the polishing monitoring value, which is a first derivative, a second derivative, or a higher derivative of the torque current measurement, may be generated.

The time-series data shown in FIG. 7 includes an initial polishing data point IP and two characteristic data points FP1 and FP2. Because the initial polishing data point IP and the two feature data points FP1 and FP2 correspond to the initial polishing data point IP and the two feature data points FP1 and FP2 in the time series data of the learning workpiece shown in FIG. , the redundant description of which will be omitted. In the example shown in FIG. 7, the polishing end point predicting device 33 generates time-series data from the initial polishing data point IP to the portion indicated by the solid line after the characteristic data point FP2 appears.

The polishing end point predicting device 33 detects the initial polishing data point IP and the feature data points FP1 and FP2 from the time-series data generated when polishing the workpiece W, and learns the polishing monitoring value and the polishing time that constitute each data point. Input to the input layer 101 (see FIG. 5) of the finished model 34 . More specifically, the polishing end point predicting device 33 calculates the polishing monitoring value VI and the polishing time TI that form the initial polishing data point IP, the polishing monitoring value VF1 and the polishing time TF1 that form the characteristic data point FP1, and FP2. The constituent polishing monitoring value VF2 and polishing time TF2 are input to the input layer 101 .

The polishing end point predicting device 33 uses the data input to the input layer 101 of the trained model 34 to perform operations according to the algorithm defined by the trained model 34, and the output layer 103 (see FIG. 5) of the trained model 34. reference) outputs the predicted polishing end point time. The polishing end point prediction device 33 sends a polishing end signal including the output polishing end point prediction time to the polishing control section 40 . Based on the received polishing end signal, the polishing control unit 40 issues a command to the polishing module 1 to end the polishing of the workpiece W at the polishing end point prediction time.

In the above-described embodiment, the data points input to the learned model 34 during polishing of the workpiece W are the initial polishing data point IP and the two feature data points FP1 and FP2. Only the two feature data points FP1 and FP2 may be used without using the point IP.

In another embodiment, as training data used to construct the learned model 34, as shown in FIG. The data points that are input to the trained model 34 during polishing may be characteristic data points at which predetermined polishing monitoring values are reached.

In still another embodiment, when the time-series data itself from the start of polishing to the end point of polishing is used as the training data used to construct the learned model 34, the time-series data from the start of polishing is used during polishing of the workpiece W. The data itself may be input to the trained model 34 and the predicted polishing end point time may be output.

According to the above-described embodiment, the polishing monitoring value and the polishing time constituting the data points including the feature data points appearing before the polishing end point are input to the learned model 34 constructed by machine learning. The end point prediction time is output from the trained model 34 . As a result, the polishing end point time can be accurately predicted before the target polishing end point, and polishing of the workpiece can be finished at the target polishing end point.

FIG. 8 is a flow chart describing one embodiment of a method for predicting the polishing endpoint of a workpiece W using a trained model 34. As shown in FIG.
In step 1 a learning workpiece is polished by polishing module 1 . The learning workpiece is polished under the same polishing conditions as the workpiece W. A polishing progress meter 30 measures the torque current during polishing of the learning workpiece.
In step 2 , the polishing end point predicting device 33 acquires from the polishing progress measuring device 30 a measured value of the torque current indicating the polishing progress of the learning workpiece, that is, a polishing monitoring value.

In step 3, the polishing endpoint predictor 33 generates time series data representing changes in polishing monitoring values (ie, changes in torque current measurements) over time for polishing the learning workpiece.
In step 4, the polishing end point prediction device 33 performs machine learning using data points included in the time series data so that an appropriate polishing end point prediction time is output. More specifically, the polishing endpoint prediction device 33 generates training data including polishing monitoring values and polishing times that form data points including feature data points appearing before the polishing endpoint, and polishing endpoint times as correct labels. machine learning using A learned model 34 is constructed by repeating the above steps 1 to 4 using a plurality of time-series data obtained when polishing a plurality of learning workpieces (step 5).

In step 6 , the workpiece W to be polished is polished by the polishing module 1 .
In step 7 , the polishing end point predicting device 33 acquires from the polishing progress measuring device 30 a torque current measurement value indicating the polishing progress of the workpiece W, that is, a polishing monitoring value.
In step 8, the polishing endpoint prediction device 33 generates time-series data representing changes in polishing monitoring values (that is, changes in measured values of torque current) along the polishing time of the workpiece W. FIG.

In step 9 , the polishing end point prediction device 33 inputs data points included in the time-series data generated during polishing of the workpiece W to the learned model 34 . More specifically, the polishing endpoint prediction device 33 inputs polishing monitoring values and polishing times that constitute data points, including feature data points appearing before the polishing endpoint, to the trained model 34 .
At step 10 , the polishing end point prediction device 33 outputs the polishing end point prediction time from the learned model 34 .
In step 11 , the polishing end point prediction device 33 sends a polishing end signal including the output polishing end point prediction time to the polishing control section 40 .
At step 12, the polishing controller 40 issues a command to the polishing module 1 based on the received polishing end signal to finish polishing the workpiece W at the polishing end point prediction time.

The polishing end point prediction device 33 operates according to instructions included in a program electrically stored in the storage device 33a. That is, when the workpiece W is polished, the polishing end point predicting device 33 acquires a polishing monitoring value indicating the polishing progress of the workpiece W, and when representing a change in the polishing monitoring value along the polishing time of the workpiece W A step of generating series data, and a step of inputting polishing monitoring values and polishing times that constitute characteristic data points included in the time series data to the trained model 34 and outputting the predicted polishing end point time from the trained model 34 are performed. do.

A program for causing the polishing endpoint predicting device 33 to execute these steps is recorded in a computer-readable recording medium, which is a non-temporary tangible object, and provided to the polishing endpoint predicting device 33 via the recording medium. Alternatively, the program may be input to the polishing endpoint prediction device 33 via a communication network such as the Internet or a local area network.

In the above-described embodiment, the data points of the training data used for building the trained model 34 and the data points input to the trained model 34 to output the polishing end point prediction time are the polishing initial Although data point IP and feature data points FP1 and FP2 are used, other embodiments may use intermediate data points described below in addition to these data points.

As shown in FIG. 9, the intermediate data points are four data points DP1, DP2, DP3, DP4 included in the time series data. Intermediate data points DP1 and DP2 are data points between the initial polishing data point IP and the feature data point FP1. Intermediate data points DP3, DP4 are data points between two feature data points FP1, FP2. The intermediate data points DP1 to DP4 are data points on time-series data corresponding to a plurality of polishing times distributed at predetermined time intervals B, for example. Although the number of intermediate data points shown in FIG. 9 is four, the number of intermediate data points is not limited to this, and may be three or less, or five or more.

In this case, the training data used to build the trained model 34 is the feature data points and intermediate data points DP1 to DP4 appearing up to the polishing end point, which are included in the time-series data generated by polishing the learning workpiece. Includes polishing monitoring value and polishing time to configure, and polishing endpoint time as correct label. The training data may further include polishing monitoring values and polishing times that constitute initial polishing data points IP for the training workpiece.

Further, the polishing end point predicting device 33 configures characteristic data points and intermediate data points DP1 to DP4 appearing until the polishing end point, which are included in the time-series data of the polishing monitoring values generated during polishing of the workpiece W to be polished. The polishing monitoring value and the polishing time are input to the learned model 34 , and the predicted polishing end point time is output from the learned model 34 . When the training data includes the initial polishing data points IP, the polishing end point predicting device 33 adds time-series data of polishing monitoring values generated during polishing of the workpiece W, in addition to the feature data points and the intermediate data points. The polishing monitoring values and polishing times that make up the included initial polishing data points IP are input into the trained model 34 .

By using the intermediate data points DP1-DP4 as training data data points used to build the trained model 34 and data points input to the trained model 34 to output the polishing endpoint prediction time, the time series Data waveforms can be specified more accurately. Therefore, the polishing end point prediction device 33 can improve the accuracy of the polishing end point prediction time output from the learned model 34 .

In yet another embodiment, the training data used to build the trained model 34 includes the total hours of use of the polishing pad 2 used in the polishing module 1, the type of polishing pad 2, the polishing of the training workpiece, and the It may further include auxiliary information of at least one of the type of polishing fluid being used and the type of learning workpiece. The polishing end point prediction device 33 executes machine learning using training data including polishing monitoring values and polishing times that form characteristic data points, the auxiliary information, and the polishing end point time as a correct label. to build. As noted above, the training data may further include polishing monitoring values and polishing times that constitute initial polishing data points and/or intermediate data points.

In this case, the polishing end point prediction device 33 configures data points including characteristic data points appearing at least up to the polishing end point, which are included in the time-series data of the polishing monitoring values generated during polishing of the workpiece W to be polished. In addition to the monitoring value and the polishing time, the auxiliary information is further input to the learned model 34, and the predicted polishing end point time is output from the learned model 34. FIG. More specifically, the polishing end point prediction device 33 determines the total usage time of the polishing pad 2 used in the polishing module 1, the type of the polishing pad 2, the type of polishing liquid used for polishing the workpiece W, and at least one of the auxiliary information of the type of workpiece W is further input to the learned model 34 , and the predicted polishing end point time is output from the learned model 34 .

The total usage time of the polishing pad 2 used in the polishing module 1, the type of polishing pad 2, the type of polishing liquid used for polishing the workpiece, and the type of workpiece, which are auxiliary information, are all It is a parameter that affects series data (ie, changes in polishing monitoring values). By inputting such auxiliary information, the polishing end point prediction device 33 can improve the accuracy of the polishing end point prediction time output from the learned model 34 . In particular, since the type of workpiece is highly related to the appearance pattern of feature data points on the time series data, by using the type of workpiece as auxiliary information, even if the number of learning workpieces is small Polishing end point time can be predicted with high accuracy.

In the embodiments described so far, the polishing monitoring value is a measurement of torque current that varies due to friction between the workpiece W and the polishing pad 2 . The frictional force generated between the workpiece W and the polishing pad 2 is generated as the workpiece W and the polishing pad 2 move relative to each other. The table rotation motor 18 is an example of a motor that relatively moves the workpiece W and the polishing pad 2 . The frictional force generated between the workpiece W and the polishing pad 2 is the torque supplied to the polishing head rotation motor 12 that rotates the polishing head 6 while the polishing head 6 is pressing the workpiece W against the polishing pad 2. It also causes a change in current. Therefore, in one embodiment, as shown in FIG. 10, the polishing progress measuring device 30 is a torque current detector (that is, a polishing head rotation torque current detector) that measures the torque current supplied to the polishing head rotation motor 12. may be Furthermore, when the polishing head 6 is oscillated by the arm oscillating motor 16 during polishing of the workpiece W, as shown in FIG. It may be a torque current detector that measures current (that is, a polishing head swing torque current detector).

Although the polishing endpoint detection system 7 described so far uses torque current measurements as polishing monitoring values, polishing monitoring values are not limited to torque current measurements. In one embodiment, as shown in FIG. 12, the polishing progress measuring device 30 is a film thickness measurement sensor that measures the film thickness of the workpiece W, and the polishing monitoring value is the film thickness index value of the workpiece W. good. In the embodiment shown in FIG. 12, the film thickness measurement sensor is arranged within the polishing table 3 and rotates together with the polishing table 3 . The film thickness measurement sensor is configured to measure the film thickness of the workpiece W while traversing the workpiece W on the polishing pad 2 . Examples of film thickness sensors include eddy current sensors and optical sensors. A known configuration can be adopted for measuring the film thickness using such a film thickness measurement sensor.

The above-described embodiments are described for the purpose of enabling a person having ordinary knowledge in the technical field to which the present invention belongs to implement the present invention. Various modifications of the above embodiments can be made by those skilled in the art, and the technical idea of the present invention can be applied to other embodiments. Accordingly, the present invention is not limited to the described embodiments, but is to be construed in its broadest scope in accordance with the technical spirit defined by the claims.

W workpiece 1 polishing module 2 polishing pad 2a polishing surface 3 polishing table 3a table axis 5 head shaft 6 polishing head 7 polishing end point detection system 10 torque transmission mechanism 12 polishing head rotation motor 15 head arm 16 arm swing motor 18 table rotation motor 20 Polishing Liquid Supply Nozzle 30 Polishing Progress Measuring Device 33 Polishing End Point Predicting Device 33a Storage Device 33b Arithmetic Device 34 Learned Model 40 Polishing Control Unit 101 Input Layer 102 Hidden Layer 103 Output Layer

Claims (20)

Acquiring a polishing monitoring value indicating the polishing progress of the workpiece when polishing the workpiece;
generating time series data representing changes in the polishing monitoring values over time for polishing the workpiece;
A polishing endpoint detection method comprising: inputting a polishing monitoring value and a polishing time that constitute feature data points included in the time-series data into a learned model; and outputting a polishing endpoint prediction time from the learned model. 2. The method according to claim 1, further inputting a polishing monitoring value and a polishing time that constitute polishing initial data points included in said time-series data into said trained model, and outputting a polishing end point prediction time from said trained model. Polishing endpoint detection method. A polishing monitoring value and a polishing time that constitute an intermediate data point between the initial polishing data point and the feature data point, which are included in the time-series data, are further input to the trained model, and from the trained model 3. The polishing endpoint detection method according to claim 2, wherein a polishing endpoint prediction time is output. 2. The polishing endpoint detection method according to claim 1, wherein said feature data points are a plurality of feature data points included in said time-series data. The polishing endpoint predicting device further inputs a polishing monitoring value and a polishing time that constitute intermediate data points between the plurality of feature data points included in the time-series data into the learned model, 5. The polishing endpoint detection method according to claim 4, wherein the polishing endpoint prediction time is output from the model. at least one auxiliary information of the total usage time of the polishing pad used in the polishing module, the type of the polishing pad, the type of polishing liquid used for polishing the workpiece, and the type of the workpiece; 6. The polishing endpoint detection method according to claim 1, further inputting to a learned model and outputting a polishing endpoint prediction time from said learned model. 7. The polishing endpoint detection method according to claim 1, wherein said characteristic data point is one of a maximum point, a minimum point and an inflection point of said time-series data. 7. The polishing endpoint detection method of claim 1, wherein the characteristic data point is a data point when a predetermined polishing monitoring value is reached. 9. The polishing endpoint detection method according to claim 1, wherein said polishing progress measuring device is one of a torque current detector and a film thickness measuring sensor. a polishing progress measuring device provided in the polishing module;
Equipped with a polishing endpoint prediction device having a learned model constructed by machine learning,
The polishing end point predicting device includes:
Acquiring a polishing monitoring value indicating the polishing progress of the workpiece output from the polishing progress measuring device during polishing of the workpiece;
generating time series data representing changes in the polishing monitoring values over time for polishing the workpiece;
Polishing end point detection configured to input a polishing monitoring value and a polishing time, which constitute feature data points included in the time-series data, into the learned model, and output a polishing end point prediction time from the learned model. system. The polishing endpoint predicting device further inputs a polishing monitoring value and a polishing time that constitute initial polishing data points included in the time-series data to the learned model, and outputs a polishing endpoint prediction time from the learned model. 11. The polishing endpoint detection system of claim 10, configured to: The polishing endpoint predicting device further inputs a polishing monitoring value and a polishing time, which constitute an intermediate data point between the initial polishing data point and the feature data point, into the learned model. 12. The polishing endpoint detection system according to claim 11, further configured to output a polishing endpoint prediction time from said learned model. 11. The polishing endpoint detection system of claim 10, wherein the feature data points are a plurality of feature data points included in the time series data. The polishing endpoint predicting device further inputs a polishing monitoring value and a polishing time that constitute intermediate data points between the plurality of feature data points included in the time-series data into the learned model, 14. The polishing endpoint detection system of claim 13, configured to output a polishing endpoint prediction time from the model. The polishing end point predicting device determines the total usage time of the polishing pad used in the polishing module, the type of the polishing pad, the type of polishing liquid used for polishing the workpiece, and the type of the workpiece. 15. The polishing according to any one of claims 10 to 14, further configured to input at least one auxiliary information among them to said trained model and output a polishing end point prediction time from said trained model. Endpoint detection system. 16. The polishing endpoint detection system according to any one of claims 10 to 15, wherein said characteristic data point is any one of a local maximum point, a local minimum point and an inflection point of said time series data. 16. The polishing endpoint detection system of any one of claims 10-15, wherein the characteristic data point is a data point at which a predetermined polishing monitoring value is reached. 18. The polishing endpoint detection system according to any one of claims 10 to 17, wherein said polishing progress measuring device is one of a torque current detector and a film thickness measuring sensor. a polishing table for supporting the polishing pad;
a polishing head for polishing the workpiece by pressing the workpiece against the polishing surface of the polishing pad;
A polishing apparatus comprising the polishing endpoint detection system according to any one of claims 10 to 18. obtaining polishing monitoring values indicative of polishing progress of the workpiece during polishing of the workpiece;
generating time series data representing changes in the polishing monitoring values over time for polishing the workpiece;
A computer storing a program for causing the computer to execute steps of inputting a polishing monitoring value and a polishing time, which constitute characteristic data points included in the time-series data, into a trained model and outputting a predicted polishing end point time from the trained model. A readable recording medium.

Images