JP4836994B2 - Semiconductor processing equipment - Google Patents

Description

  The present invention relates to a semiconductor processing apparatus, and more particularly to a semiconductor processing apparatus that predicts a processing result and improves the operation rate and reliability of the apparatus.

  In recent years, the dimensions of semiconductor devices have been miniaturized, and the dimensional accuracy of processing has become so severe that a gate electrode of 0.1 μm or less must be processed with a dimensional accuracy of 10% or less. On the other hand, in semiconductor manufacturing equipment that physicochemically processes semiconductor wafers using heat or plasma, reaction products generated by chemical reactions inside the equipment remain attached to the inner wall of the equipment, and over time, Change the processing state. For this reason, as the number of wafer processes increases, the processing shape of the semiconductor device on the wafer gradually changes and the performance deteriorates.

  For this problem, usually, measures are taken such as cleaning the deposits on the inner wall of the chamber with plasma, or increasing the temperature of the chamber walls to make the deposits difficult to adhere. However, in most cases, these measures are not perfect, and as a result, the processing shape of the semiconductor device gradually changes. For this reason, when the machining shape becomes a problem, the parts of the manufacturing apparatus must be replaced or cleaned before the machining shape is changed. In addition to the deposited film, various apparatus state variations are involved in wafer shape changes.

  For this reason, a device has been devised such as detecting a change in the processing state inside the semiconductor manufacturing apparatus and feeding back the detection result to the input of the semiconductor manufacturing processing apparatus to keep the processing state constant.

  A method for monitoring fluctuations in plasma processing is disclosed in Patent Document 1, for example. This publication discloses a method for predicting apparatus performance and diagnosing the state of plasma using a relational expression between plasma processing characteristics and apparatus electrical signals. As the method, a method is disclosed in which an approximate expression representing the relationship between three electrical signals and the plasma processing characteristics of the apparatus is obtained by multiple regression analysis.

  Another example is shown in Patent Document 2. This publication discloses a method of applying a general detection system having a plurality of existing detectors to a semiconductor manufacturing apparatus and monitoring the state of the apparatus from a correlation signal of the detection signals. As a method for generating the correlation signal, a calculation formula based on a ratio of six electrical signals is disclosed.

Another example is shown in Patent Document 3. This publication discloses a method for monitoring the state of an apparatus by taking a large number of signals from light and mass analyzers to generate a correlation signal. As a method for generating the correlation signal, a method using principal component analysis is shown.
JP-A-10-125660 Japanese Patent Laid-Open No. 11-87323 US Pat. No. 5,658,423

 However, in the method of Patent Document 1, when there are many types of sensor data for monitoring the apparatus, many signals unrelated to the processing performance to be predicted are included in the explanatory variables, so the prediction by the multiple regression analysis is not successful. Disappear. The method of Patent Document 2 is a general method in which a signal obtained by correlating a plurality of detection signals from a plurality of well-known detection means is used for diagnosis. Also, the disclosed correlation method is a conventional method in which several signal ratios are taken, and these methods are used to accurately determine the state of a semiconductor manufacturing apparatus that takes various states depending on many causes of variation. It is difficult to apply to a monitoring system.

  Unlike the above method, Patent Document 3 discloses a method of monitoring the plasma state by capturing principal component analysis of a large amount of data monitored from the apparatus and capturing changes in the apparatus state. However, a semiconductor manufacturing apparatus used for actual mass production does not work well just by applying such a general statistical processing method. For example, it is almost impossible to know how the main component changes and the processing result will change.

  The present invention has been made in view of these problems. In a semiconductor processing apparatus that processes various types of devices, the processing state is monitored, and abnormal processing is detected or a processing result is predicted based on the monitor output. The present invention provides a semiconductor processing apparatus with improved operating rate and reliability of the apparatus.

  In order to solve the above problems, the present invention employs the following means.

A plurality of sensors for monitoring the process state of a semiconductor processing apparatus for processing a semiconductor wafer, a process forming HateIri force means for inputting a measured value of the processing result of the semiconductor wafer treated by the semiconductor processing apparatus, wherein the plurality of sensors model expression generation unit and for generating a model equation for predicting processing results using the acquired sensor data and the measured value as an explanatory variable to explain the variation of the processing results calculated values of said plurality of cell Nsadeta based A model formula storage unit that stores the generated model formula, a model formula stored in the model formula storage unit and a processing result prediction unit that predicts a processing result based on sensor data newly acquired by the sensor, and a processing condition control section that controls the process conditions of the semiconductor processing device so as to correct the deviation by comparing a set value set in advance and the processing result of the prediction, prior to The processing condition control unit includes processing condition control direction storage means for storing a processing condition control direction having a strong correlation with a processing result measurement value obtained by using a PLS method (Partial Least Square Method), and for correcting the deviation. The processing condition is controlled according to the corresponding processing condition control direction .

  According to the present invention, the processing state is monitored, and the abnormal processing is detected or the processing result is predicted based on the monitor output, so that the operating rate and reliability of the processing apparatus can be improved.

  FIG. 1 is a diagram showing a semiconductor processing apparatus according to the first embodiment of the present invention. In the figure, the semiconductor processing apparatus 1 is provided with a processing state monitoring unit 2 for monitoring the processing state. The processing state monitoring unit 2 may be incorporated in the semiconductor processing apparatus 1 or installed outside the processing apparatus 1. Moreover, you may install in the place distant via a network etc. Furthermore, as shown in FIG. 2, some of the functions may be separated via a network or the like.

  Details of the processing state monitoring unit 2 are shown below. First, the sensor 3 for monitoring the processing state of the wafer in the processing apparatus 1 is provided. The sensor 3 usually uses several types of sensors. For example, when the processing apparatus is a plasma etching apparatus, a plasma CVD apparatus, or the like, the sensor 3 spectrally decomposes the light emission of the plasma being processed using a spectroscope, and obtains the emission intensity of each resolved wavelength as sensor data. To do. For example, when a spectrometer having a 1000-channel CCD array is used, 1000 sensor data can be acquired for each sampling. Further, the pressure, temperature, gas flow rate, etc. of the apparatus are also used as sensor data. Also, electrical measurement results such as current, voltage, impedance, and their harmonic components can be used as sensor data.

  During the processing of the wafer, these sensor data are acquired at appropriate time intervals, and the acquired sensor data is stored in the sensor data storage unit 4. On the other hand, the processing result of the processed wafer is measured outside the processing apparatus 1 or by a processing result measuring instrument incorporated in the apparatus. The processing result measurement is measurement of a gate width by CDSEM, measurement of a processed shape such as a cross-sectional shape by a cross-sectional SEM, or measurement of electrical characteristics of a processed device. These measurements do not need to be performed for all wafers, and it is usually sufficient to extract a part of the wafers and measure the processing results.

  The processing state monitoring unit 2 has processing result measurement value input means 5 for receiving the measurement value of the processing result. The input means 5 can be a reading device that reads information recorded on a portable medium such as a flexible disk or a CDROM. Also, a wired or wireless network connection device can be used.

  The processing result measurement value received by the input means 5 is stored in the processing result measurement value storage unit 6. The storage unit 6 stores the measurement value of the processing result for each type of device.

  The model formula generation unit 7 extracts samples of the same type of device in which the sensor data and the processing result measurement value data are stored from the sensor data storage unit 4 and the processing result measurement value storage unit 6. For example, when the number of samples is 3 or more, a model formula for predicting a processing result measurement value can be created using sensor data as an explanatory variable. Usually, there are a large number of types of sensors and sensor data at this time, and it is difficult to automatically select sensors and sensor data used for prediction. In particular, when various devices are processed, the type of sensor data that is effective for prediction differs from device to device. Therefore, it is difficult to determine a sensor to be used for prediction in advance.

  FIG. 6 is a diagram for explaining model formula generation processing by the PLS method. As shown in FIG. 6, in the PLS method, explanatory variables that have the strongest correlation with fluctuations in data to be predicted are automatically generated from a large number of sensor data. At this time, a function for calculating the explanatory variable from the sensor data is also obtained.

  First, the processing result measurement value of n wafers is set as a prediction target, and the processing result measurement value of the i-th wafer is represented by Yi. When m pieces of sensor data are obtained from one wafer, the j-th sensor data of the i-th wafer is indicated by Sij. In this case, the m pieces of sensor data may be data at different times of the same sensor, or may be data from different sensors.

  FIG. 7 is a diagram illustrating an example of sensor data Sij. As shown in FIG. 7, the processing performed on one wafer 1 in the processing apparatus 1 has three steps (step 1 to step 3). At this time, sensors for monitoring the processing state are A, B, and C. Are obtained as S11 to Sn9 as shown in the figure. Sij may be an average value of sensor data during each step process, or may be a value obtained by converting sensor data such as a square or an inverse number.

  When the PLS method is used, a plurality of sensor data Sij can be converted into m explanatory variables Xik arranged in the order of the strength of correlation with the fluctuation of the wafer processing result measurement value Yi. A function Fk for converting the sensor data Sij into the explanatory variable Xk is expressed by Expression (1).

Xik = Fk (Si1, Si2,..., Sim) (1)
A processing result measurement value is predicted using some of the explanatory variables Xik. Normally, since the explanatory variable Xi1 has the strongest correlation with the processing result measurement value Yi, Xi1, Xi2, Xi3, etc. are selected as explanatory variables. In the PLS method, a prediction formula like Formula (2) is generated simultaneously. However, it may be better to generate the prediction formula (2) using the explanatory variables such as Xi1 described above.

Yi = p (Xi1, Xi2, Xi3) (2)
By the way, the processing result measurement value includes abnormal data of a wafer which has been processed abnormally due to a bad processing state of the wafer. When prediction is performed by normal multiple regression analysis including such data, a model formula with poor prediction accuracy is generated due to the abnormal data.

  FIG. 8 is a graph for explaining the robust regression analysis. When the robust regression analysis is used for prediction, as shown in FIG. 8, abnormal data is excluded from the prediction target as an outlier. Therefore, a correct prediction model formula can be generated.

  FIG. 9 is a flowchart for explaining the model formula creation process of the model formula generation unit 7. When there are many types of sensor data in the model formula generation unit 7, principal component analysis of the sensor data is performed (steps 901 and 902), and robust regression analysis is performed using the obtained first principal component to predict the processing result. (Steps 905 to 906). At this time, since the principal components that are not necessary for the prediction of the processing result are included in the explanatory variables, the principal components having a small regression coefficient are removed (step 907), and another principal component (second principal component) is used as the explanatory variable. The process of adding (step 904) and performing multiple regression analysis again (step 906) is repeated until the prediction error becomes smaller than the set value (step 908). When the prediction error becomes smaller than the set value, the process is terminated (step 909). These regression analyzes may be linear, or a non-linear regression analysis derived from the physical characteristics of the processing and experience values may be used.

  The model formula generated by such a method is stored in the model formula storage unit 8 shown in FIG. Since model formulas are created for each device of the same type, as many model formulas as the number of devices processed by the processing apparatus 1 are stored in the model formula storage unit.

  When processing by the processing apparatus 1 is started and a wafer on which a specific device is to be formed is loaded and processed in the processing apparatus 1, the prediction unit 9 loads a model formula corresponding to the specific device from the model formula storage unit 8. To do. A signal obtained from the sensor 3 during the processing of the wafer is converted into an explanatory variable using, for example, Expression 1 obtained by the PLS method, or converted into a principal component by principal component analysis. Is used to calculate the predicted value of the processing result. The calculated predicted value is passed to the processing condition control unit 10. The processing condition control unit 10 changes the processing condition so as to correct the deviation between the predicted value and the set value of the processing result.

  Next, correction of processing conditions by the processing condition control unit 10 will be described. Here, the PLS method is used again. In the processing of normal semiconductor devices, several conflicting processing performances are often required as processing requirements. For example, in the etching processing of the gate electrode, the verticality of the side wall of the gate electrode, the base oxide film, and the etching selectivity of the gate polysilicon are required.

  That is, in order to improve the verticality of the sidewall, it is better to use etching conditions with low deposition properties, and in order to achieve high selectivity with the base oxide film, it is better to use etching conditions with high deposition properties. Thus, when there are two conflicting requirements, it is difficult to control the processing conditions. 10, 11, and 12 are diagrams for explaining a method for obtaining processing conditions that satisfy such conflicting requirements.

  For example, when the verticality of the sidewall is deteriorated due to the change of the processing apparatus 1 over time, the selectivity of the base oxide film is simultaneously improved even if the verticality is improved by reducing the processing condition 1 (here, the flow rate of the gas A). Is worse as the processing conditions.

  For this reason, for example, it is necessary to combine the processing condition 1 and the processing condition 2 (wafer bias power in this case) to find a condition that improves the verticality of the side wall and does not deteriorate the selectivity of the base oxide film.

  For this purpose, first, as shown in FIG. 10, processing conditions are set by setting experimental conditions in which several to several tens of processing conditions are changed around normal processing conditions (central conditions), and processing results are obtained. Measure. Points 1 to 4 in FIG. 12 correspond to the experimental conditions 1 to 4 in FIG. Here, the measured value of sidewall verticality is taken as the processing result measurement value A, and the underlying oxide film selectivity is taken as the processing result measurement value B.

  Next, as shown in FIG. 11, the PLS method is applied to the above experiment, and the correlation between the two types of processing conditions and the two types of processing result measured values is examined. Then, as shown in FIG. 12, a direction A having a strong correlation with the side wall perpendicularity is obtained. Similarly, a direction B perpendicular to the direction A and having a strong correlation with the base oxide film selection ratio can be calculated from a direction having a strong correlation with the base oxide film selection ratio obtained by the PLS method. The condition direction A and the condition direction B are set in the processing condition control unit 10 shown in FIG. With this setting, when the prediction unit 9 predicts that the verticality of the sidewall is deteriorated based on the model formula, if the processing condition is changed to the condition direction A, the selection ratio of the underlying oxide film is sacrificed. Therefore, only the side wall perpendicularity can be improved. The control direction of the calculated processing condition is stored in the processing condition control direction storage unit 14 of the control condition control unit 10 to correct the processing condition when the processing result predicted value based on the model formula deviates from the set value. Use.

  In the above example, two types of processing conditions are changed. However, in the PLS method, more types of processing conditions can be changed, and a more favorable result is obtained as more processing conditions are changed. In addition, not only two types of conflicting processing result measurement values but also a larger number of processing results can be targeted. For example, in addition to the verticality of the side wall and the base oxide film selection ratio, the mask selection ratio can be targeted.

  FIG. 3 is a diagram showing another embodiment of the present invention. In the figure, 11 is a predicted value display unit, which displays a predicted value or a deviation between the predicted value and a preset value. In the figure, the same parts as those shown in FIG. The display unit can be a sending means such as a buzzer for issuing an alarm or an e-mail.

  FIG. 4 is a diagram showing still another embodiment of the present invention. In the above description, the monitoring control of the processing condition control unit 10 has been described on the assumption that a model formula has been generated. However, it is not possible to monitor and control a type of device for which a model formula has not yet been generated. In many cases, the measurement of the processing result is very time-consuming, and the measurement of the processing result is rarely performed. Therefore, a model formula cannot be generated for such a device. FIG. 4 is a diagram showing a processing state monitoring unit (sub processing state monitoring unit) 2 ′ that can be monitored and controlled even in such a case.

  As shown in FIG. 4, the principal component extraction unit 12 extracts principal components based on various sensor data from a large number of sensors 3. The abnormality monitoring unit 13 detects a processing abnormality by monitoring the variation of the extracted principal component fluctuations. If an abnormality is detected, the start of the next wafer processing may be stopped. For the detection of abnormality, for example, a variation management method called SPC (Statistical Process Control) may be used. For this purpose, the average value and variance of the main component being processed by the corresponding device are stored, and the process is determined to be abnormal when the measured main component is more than several times the variance from the average value.

  FIG. 5 is a diagram illustrating a processing flow suitable for a processing apparatus including both the processing state monitoring unit 2 and the sub processing state monitoring unit 2. First, it is determined whether or not a model formula for the device to be processed is generated and stored (step 501). If the model formula is stored, the process state monitoring unit 2 executes monitoring control (step 502). If the model formula is not stored, monitoring control is executed by the processing state monitoring unit 2 '(step 503).

It is a figure which shows the semiconductor processing apparatus concerning embodiment of this invention. It is a figure which shows the modification of a semiconductor processing apparatus. It is a figure which shows other embodiment of this invention. It is a figure which shows other embodiment of this invention. It is a figure explaining the process flow suitable for the processing apparatus provided with the process state monitoring part and the sub process state monitoring part. It is a figure explaining the model formula production | generation process by PLS method. It is a figure which shows the example of sensor data. It is a graph explaining a robust regression analysis. It is a flowchart explaining the model formula creation process of a model formula production | generation part. It is a figure explaining the method of calculating | requiring process conditions. It is a figure explaining the method of calculating | requiring process conditions. It is a figure explaining the method of calculating | requiring process conditions.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Semiconductor processing apparatus 2 Process state monitoring part 3 Sensor 4 Sensor data preservation | save part 5 Process result measured value input means 6 Process result measurement value preservation | save part 7 Model formula production | generation part 8 Model formula preservation | save part 9 Prediction part by model formula 10 Processing condition control Unit 11 Predicted value display unit 12 Principal component extraction unit 13 Abnormality detection unit 14 Processing condition control direction storage means

Claims (5)

A plurality of sensors for monitoring a processing state of a semiconductor processing apparatus for processing a semiconductor wafer;
A process forming HateIri force means for inputting a measured value of the processing result of the semiconductor wafer treated by the semiconductor processing apparatus,
Generating a model equation for predicting processing results is used as explanatory variables to explain the variation of the processing results calculated values of said plurality of cell Nsadeta based on the sensor data and the measured values of the plurality of sensors is acquired A model formula generator,
A model formula storage unit for storing the generated model formula ;
A processing result prediction unit that predicts a processing result based on the model formula stored in the model formula storage unit and the sensor data newly acquired by the sensor;
And a processing condition control section that controls the process conditions of the semiconductor processing device so as to correct the deviation by comparing a setting value set in advance and the processing result of the above prediction,
The processing condition control unit includes processing condition control direction storing means for storing a processing condition control direction having a strong correlation with a processing result measurement value obtained using a PLS method (Partial Least Square Method), and correcting the deviation. The semiconductor processing apparatus controls the processing conditions according to the processing condition control direction corresponding to the above . 2. The semiconductor according to claim 1, wherein the processing condition control direction storage means stores a direction of a condition having a strong correlation with the side wall perpendicularity and a direction having a strong correlation with the base oxide film selectivity. Processing equipment. The semiconductor processing apparatus according to claim 1, wherein the model formula generation unit generates a model formula using a PLS method (Partial Least Square Method) . The semiconductor processing apparatus according to claim 1, wherein the model formula generation unit generates a model formula using a robust regression analysis method . The semiconductor processing apparatus according to claim 1, wherein the model formula generation unit generates a model formula using a principal component robust regression analysis method (Principal Component Robust Regression).

Images