Semiconductor process prediction method and device considering overall characteristics and local characteristics

Abstract

A semiconductor process prediction method and device that considers overall characteristics and local characteristics. The semiconductor process prediction method includes the following steps. Obtain a plurality of machine sensing curves. Screen these machine sensing curves to reduce the collinearity of these machine sensing curves. Correct these machine sensing curves with a dynamic time correction program. Input these corrected machine sensing curves into a convolutional neural network model to obtain a first prediction result that considers local characteristics. Perform a statistical analysis program on these machine sensing curves to obtain a plurality of statistical data. Input these statistical data into an artificial neural network model to obtain a second prediction result that considers overall characteristics. Obtain a total prediction result based on the first prediction result and the second prediction result.

Description

Technical Field

[0001] The present invention relates to a semiconductor process prediction method and device, and in particular to a semiconductor process prediction method and device taking into account overall characteristics and local characteristics. Background Art

[0002] With the development of semiconductor technology, various complex semiconductor products are constantly being introduced. In the semiconductor process, wafers need to go through thousands of processes to produce the final product. Therefore, researchers must use appropriate prediction methods for semiconductor processes to predict the occurrence of process anomalies to avoid a large number of defective products in the final product.

[0003] Traditionally, statistical data such as average values ​​or standard deviations are monitored through various process values ​​to estimate possible process anomalies. However, this approach only considers overall characteristics, and with the increasing complexity of semiconductor processes, it is difficult to obtain highly accurate prediction results. Summary of the invention

[0004] The present invention relates to a semiconductor process prediction method and device that considers overall characteristics and local characteristics, which utilizes a dynamic time correction program, a convolutional neural network model and an artificial neural network model to analyze local characteristics and overall characteristics to improve prediction accuracy.

[0005] According to a first aspect of the present invention, a semiconductor process prediction method considering overall characteristics and local characteristics is proposed. The semiconductor process prediction method includes the following steps. Obtain a plurality of machine sensing curves. Screen these machine sensing curves to reduce the collinearity of these machine sensing curves. Correct these machine sensing curves using a dynamic time warping (DTW) procedure. Input these corrected machine sensing curves into a convolutional neural network model (CNN model) to obtain a first prediction result considering local characteristics. Perform a statistical analysis procedure on these machine sensing curves to obtain a plurality of statistical data. Input these statistical data into an artificial neural network model (ANN model) to obtain a second prediction result considering overall characteristics. Obtain a total prediction result based on the first prediction result and the second prediction result.

[0006] According to a first aspect of the present invention, a semiconductor process prediction device considering overall characteristics and local characteristics is provided. The semiconductor process prediction device includes a database, a screening unit, a correction unit, a convolutional neural network model (CNN model), a statistical unit, an artificial neural network model (ANN model) and a total prediction unit. The database is used to store a plurality of machine sensing curves. The screening unit is used to screen the machine sensing curves to reduce the collinearity of the machine sensing curves. The correction unit corrects the machine sensing curves by a dynamic time warping (DTW) procedure. The convolutional neural network model is used to receive the corrected machine sensing curves to obtain a first prediction result considering local characteristics. The statistical unit is used to perform a statistical analysis procedure on the machine sensing curves to obtain a plurality of statistical data. The artificial neural network model is used to receive the statistical data to obtain a second prediction result considering overall characteristics. The total prediction unit is used to obtain a total prediction result based on the first prediction result and the second prediction result.

[0007] In order to better understand the above and other aspects of the present invention, embodiments are given below and described in detail with reference to the accompanying drawings: BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 A semiconductor process prediction device according to an embodiment is shown.

[0009] Figure 2 A block diagram of a semiconductor process prediction device according to an embodiment is shown.

[0010] Figure 3 A flow chart of a semiconductor process prediction method considering both global characteristics and local characteristics according to an embodiment is shown.

[0011] Figure 4 Example Figure 3 each step.

[0012] Figure 5 A flowchart showing detailed steps of step S120 is shown.

[0013] Figure 6 Example Figure 5 each step.

[0014] Figure 7 An implementation of a convolutional neural network model is shown.

[0015] Figure 8Another implementation of a convolutional neural network model is shown. DETAILED DESCRIPTION

[0016] Please refer to Figure 1 , which illustrates a semiconductor process prediction device 100 according to an embodiment. In the semiconductor process, the wafer needs to pass through various machines 900 to perform different processes, such as deposition, etching, or annealing. Each process of each machine 900 requires precise monitoring. Therefore, various sensors 910 are set on the machine 900 to extract sensing data S01~S06 such as temperature, pressure, and gas concentration (the number is not intended to limit the present invention). These sensing data S01 are continuously extracted over time and transmitted to the semiconductor process prediction device 100 through the network 800 to be recorded as a machine sensing curve S11 (shown in Figure 2 ); These sensing data S02 are continuously extracted over time and recorded as a machine sensing curve S12 (shown in Figure 2 ), and so on. With these machine sensing curves S11-S16 (the number is not intended to limit the present invention), the semiconductor process prediction device 100 can perform fault detection and classification (FDC). When predicting abnormal process conditions, the machine can be shut down or corrected immediately to avoid process defects.

[0017] Please refer to Figure 2, which illustrates a block diagram of a semiconductor process prediction device 100 according to an embodiment. The semiconductor process prediction device 100 is, for example, a server, a computer or a cloud computing center. The semiconductor process prediction device 100 includes a database 110, a screening unit 120, a correction unit 130, a convolutional neural network model (CNN model) 140, a statistics unit 150, an artificial neural network model (ANN model) 160 and a total prediction unit 170. The functions of each component are summarized as follows. The database 110 is used to store data, such as a memory, a hard disk or a cloud storage center. The screening unit 120 is used to perform data screening. The correction unit 130 is used to perform data correction. The convolutional neural network model 140, the artificial neural network model 160 and the total prediction unit 170 are used to perform data prediction. The statistics unit 150 is used to perform data statistics. The screening unit 120, the correction unit 130, the convolutional neural network model 140, the statistical unit 150, the artificial neural network model 160 and / or the total prediction unit 170 are, for example, a program code, a circuit, a chip, a circuit board or a storage device storing program code. In the present embodiment, the semiconductor process prediction device 100 can perform time correction of the curve data through the correction unit 130, and directly analyze the curve data through the convolutional neural network model 140 to consider local features. In addition, the semiconductor process prediction device 100 further performs statistics of the curve data through the statistical unit 150, and analyzes the statistical data through the artificial neural network model 160 to consider overall features. In other words, the semiconductor process prediction device 100 can consider local features and overall features at the same time to improve prediction accuracy. The operation of the semiconductor process prediction device 100 is described in detail below through a flow chart.

[0018] Please refer to Figure 3 and Figure 4 , Figure 3 A flowchart of a semiconductor process prediction method considering both global characteristics and local characteristics according to an embodiment is shown. Figure 4 Example Figure 3 In step S110, a plurality of machine sensing curves S11-S16 are obtained from the database 110. Each machine sensing curve S11-S16 is sensing data S01-S06 continuously extracted over time.

[0019] Next, in step S120, the screening unit 120 screens these machine sensing curves S11~S16 to reduce the collinearity of these machine sensing curves S11~S16. For example, a temperature increase will cause a pressure increase; a temperature decrease will also cause a pressure decrease. Therefore, there is collinearity between the temperature factor and the pressure factor, and they are essentially the same factor. If both the temperature sensing curve and the pressure sensing curve are included in the subsequent analysis, the learning and prediction of the convolutional neural network model 140 will be overly biased towards the same factor, thereby reducing the accuracy. Therefore, by reducing the collinearity of the machine sensing curve S1 through appropriate screening steps, the accuracy of the prediction can be ensured.

[0020] Please refer to Figure 5 and Figure 6 , Figure 5 A flowchart showing the detailed steps of step S120 is shown. Figure 6 Example Figure 5 Step S120 includes steps S121 and S122. Figure 6 In the example shown, there are 6 machine sensing curves S11-S16. In step S121, the screening unit 120 classifies the machine sensing curves S11-S16 into a plurality of groups G1-G3 according to a correlation matrix MX. The correlation matrix MX records the correlation coefficients between the machine sensing curves S11-S16 (e.g. Figure 6 Those with a relationship coefficient greater than a predetermined threshold are classified into the same group. Figure 6 As shown, the machine sensing curves S11-S13 are classified into group G1; the machine sensing curves S14-S15 are classified into group G2; and the machine sensing curve S16 forms a group G3.

[0021] Next, in step S122, the screening unit 120 selects one machine sensing curve (i.e., machine sensing curves S11, S15, S16) from each group G1, G2, G3. Only one machine sensing curve is selected from each group G1, G2, G3. In group G1, the correlation coefficient between the machine sensing curves S11-S13 and the predicted target Y0 (e.g., Figure 6 The one with the largest correlation coefficient (i.e., the machine sensing curve S11) can be selected. In group G2, the correlation coefficient between the machine sensing curves S14-S15 and the predicted target Y0 (as shown in FIG. Figure 6 The one with the largest value (i.e., the machine sensing curve S15) can be selected. Therefore, the selected machine sensing curves S11, S15, and S16 have low correlations with each other and do not have collinearity. In addition, the selected machine sensing curves S11, S15, and S16 have a high correlation coefficient with respect to the predicted target Y0 and are the most representative.

[0022] Then, in Figure 3 In step S130, the calibration unit 130 calibrates the machine sensing curves S11, S15, and S16 using a dynamic time warping (DTW) procedure. Figure 4 As shown, the machine sensing curve S11 will be compared with the template curve S11', and the machine sensing curve S11 and the template curve S11' can be effectively aligned through the alignment point. Similarly, the other machine sensing curves S15 and S16 will also undergo a dynamic time calibration procedure.

[0023] Then, in step S140, the calibrated machine sensing curves S11, S15, S16 are input to the convolutional neural network model 140 to obtain a first prediction result R1 considering local features. The convolutional neural network model 140 is, for example, a LeNet model, an AlexNet model, a VGG model, a GoogLeNet model, or a ResNet model.

[0024] The data input in this step is a continuous curve, and the detailed features on the continuous curve can be taken into account, including surges, drifts, oscillations, etc.

[0025] Please refer to Figure 7 , which illustrates an implementation of the convolutional neural network model 140. The convolutional neural network model 140 is, for example, a single-channel model. The single-channel model analyzes only one factor at a time, thereby avoiding interference from other factors.

[0026] Please refer to Figure 8 , which illustrates another embodiment of the convolutional neural network model 140. The convolutional neural network model 140 is, for example, a multi-channel model. The multi-channel model can analyze multiple factors at the same time and can take into account the interaction relationship of all factors at the same time.

[0027] Then, in Figure 2 In step S150, the statistical unit 150 performs a statistical analysis procedure on the machine sensing curves S11, S15, S16 to obtain a plurality of statistical data ST1, ST5, ST6. The statistical data ST1, ST5, ST6 are, for example, average values, standard deviations, medians, and the like.

[0028] Next, in step S160, these statistical data ST1, ST5, ST6 are input to the artificial neural network model 160 to obtain a second prediction result R2 that takes into account the overall characteristics. The artificial neural network model 160 is, for example, a supervised learning network, an unsupervised learning network, a hybrid learning network, an associate learning network, an optimization application network, etc. The data input in this step are statistical values ​​such as the mean, standard deviation, and median, which can take into account the overall characteristics of the continuous curve, including overall offset, overall stability, etc. The above steps S150 and S160 can be executed before step S140. Alternatively, step S140 and step S150 can be executed simultaneously.

[0029] Then, in step S170, the total prediction unit 170 obtains a total prediction result RS according to the first prediction result R1 and the second prediction result R2. In this step, the total prediction unit 170 can obtain the total prediction result RS through a voting process.

[0030] According to the above embodiment, the semiconductor process prediction device 100 and the semiconductor process prediction method can perform time correction of the curve data, and directly analyze the curve data through the convolutional neural network model 140 to consider local features. In addition, the semiconductor process prediction device 100 and the semiconductor process prediction method further perform statistics on the curve data, and analyze the statistical data through the artificial neural network model 160 to consider overall features. In other words, the semiconductor process prediction device 100 can consider both local features and overall features to improve prediction accuracy.

[0031] In summary, although the present invention has been disclosed in the above embodiments, it is not intended to limit the present invention. A person skilled in the art of the present invention may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be determined by the definition of the attached claims.

Claims

1. A semiconductor process prediction method considering overall characteristics and local characteristics, comprising: Obtain multiple machine sensing curves; Screening the machine sensing curves to reduce the collinearity of the machine sensing curves; Correcting the sensing curves of these machines with reduced collinearity using a dynamic time warping (DTW) procedure; Inputting the machine sensing curves that have been subjected to the dynamic time correction process into a convolutional neural network model (CNN model) to obtain a first prediction result that takes local features into consideration; Performing statistical analysis procedures on these machine sensing curves to obtain multiple statistical data; Inputting these statistical data into an artificial neural network model (ANN model) to obtain a second prediction result considering the overall characteristics; and Obtaining an overall prediction result according to the first prediction result considering the local features and the second prediction result considering the overall features; The sensing curves of the machines in the same batch are input into the convolutional neural network model and the artificial neural network model to obtain the first prediction result considering the local features and the second prediction result considering the overall features for the sensing curves of the machines in the same batch. 2 . The semiconductor process prediction method considering global characteristics and local characteristics as claimed in claim 1 , wherein each of the machine sensing curves is sensing data continuously extracted over time.

3. The semiconductor process prediction method considering global characteristics and local characteristics as claimed in claim 1, wherein the step of screening the tool sensing curves comprises: According to the correlation matrix, these machine sensing curves are classified into multiple groups; as well as A sensing curve of the machine is selected from each of the groups.

4. The semiconductor process prediction method considering global features and local features as described in claim 1, wherein the convolutional neural network model is a single-channel model.

5. The semiconductor process prediction method considering global features and local features as described in claim 1, wherein the convolutional neural network model is a multi-channel model. 6 . The semiconductor process prediction method considering global characteristics and local characteristics as claimed in claim 1 , wherein the step of obtaining the total prediction result according to the first prediction result and the second prediction result is to obtain the total prediction result through a voting procedure.

7. A semiconductor process prediction device considering overall characteristics and local characteristics, comprising: A database for storing sensing curves of multiple machines; A screening unit, used for screening the machine sensing curves to reduce the collinearity of the machine sensing curves; A calibration unit, which calibrates the sensing curves of the machines with reduced collinearity by a dynamic time warping (DTW) procedure; A convolutional neural network model (CNN model) is used to receive the machine sensing curves that have been subjected to a dynamic time correction process to obtain a first prediction result that takes local features into consideration; A statistical unit is used to perform a statistical analysis procedure on the sensing curves of these machines to obtain a plurality of statistical data; An artificial neural network model (ANN model) is used to receive the statistical data to obtain a second prediction result considering the overall characteristics; and A total prediction unit, used for obtaining a total prediction result according to the first prediction result considering the local features and the second prediction result considering the overall features; The sensing curves of the machines in the same batch are input into the convolutional neural network model and the artificial neural network model to obtain the first prediction result considering the local features and the second prediction result considering the overall features for the sensing curves of the machines in the same batch. 8 . The semiconductor process prediction device considering global characteristics and local characteristics as claimed in claim 7 , wherein each of the tool sensing curves is sensing data continuously extracted over time.

9. The semiconductor process prediction device considering global characteristics and local characteristics as claimed in claim 7, wherein the screening unit classifies the tool sensing curves into a plurality of groups according to a correlation matrix, and selects one tool sensing curve from each group.

10. The semiconductor process prediction device considering global features and local features as described in claim 7, wherein the convolutional neural network model is a single-channel model.

11. The semiconductor process prediction device considering global features and local features as described in claim 7, wherein the convolutional neural network model is a multi-channel model. 12 . The semiconductor process prediction device considering global characteristics and local characteristics as claimed in claim 7 , wherein the overall prediction unit obtains the overall prediction result through a voting procedure.

Images