LSTM-Based Virtual Metrology Device for Semiconductor Manufacturing Machines

Abstract

The present invention discloses a virtual metrology device for semiconductor manufacturing tools based on LSTM, comprising: an LSTM module for outputting a predicted value of metrology data according to the input data of semiconductor manufacturing tools; a window moving LSTM module for capturing the long-term drift information of process variables and metrology data of semiconductor manufacturing tools to update the model parameters of the LSTM module. The present invention can automatically predict some key metrology parameters according to the production tool data and improve the stability of prediction accuracy. For example, it can have stable prediction accuracy in different process chambers and at different times, can timely detect abnormalities in the manufacturing process, save metrology benefits, and achieve batch control.

Description

Technical Field

[0001] The present invention relates to a semiconductor integrated circuit manufacturing machine tool, and particularly to a semiconductor manufacturing machine tool virtual metrology device based on a Long Short Term Memory (LSTM) artificial neural network. Background Art

[0002] Virtual metrology, in cases where actual measurement is not possible or measurement resources need to be saved, uses production machine tool parameters and measurement parameter data to build a model and estimate the key index results of the wafers, so as to achieve real-time product quality prediction of the wafers, machine tool efficiency monitoring, and production process improvement; in this way, abnormalities can be detected immediately to avoid major losses.

[0003] Existing virtual metrology devices have the following problems:

[0004] Prediction ability: A key requirement is the accuracy of virtual metrology, that is, the prediction accuracy. The prediction ability of existing virtual metrology devices needs to be further improved.

[0005] Stability: The prediction accuracy should be maintained stable. However, even slight parameter changes and drifts during the point selection process, as well as the slow drift characteristics occurring over time during the manufacturing process, will limit the prediction accuracy of the VM model. The stability of the prediction accuracy of existing virtual metrology devices needs to be further improved.

[0006] Existing measurement data is not effectively utilized, and only based on the machine tool sensor data, the prediction accuracy is relatively low. Existing virtual metrology devices need to further improve the utilization of existing measurement data to improve the prediction accuracy. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a semiconductor manufacturing machine tool virtual metrology device based on LSTM, which can automatically predict some key measurement parameters according to the production machine tool data and improve the stability of the prediction accuracy, such as having stable prediction accuracy in different process chambers and different periods, being able to detect abnormalities in the manufacturing process in a timely manner, saving measurement benefits, and achieving run-to-run control.

[0008] To solve the above technical problem, the semiconductor manufacturing machine tool virtual metrology device based on LSTM provided by the present invention includes:

[0009] An LSTM module for outputting a predicted value of measurement data according to the input data of the semiconductor manufacturing machine tool; the predicted value of measurement data is the predicted value of the measurement data of the wafer that has completed the manufacturing process of the semiconductor manufacturing machine tool.

[0010] A window moving LSTM module is used to capture the long-term drift information of the process variables and measurement data of the semiconductor manufacturing tool, so as to update the model parameters of the LSTM module, thereby improving the stability of the prediction accuracy of the LSTM module.

[0011] A further improvement is that the initial model parameters of the LSTM module are obtained by learning the data of the known semiconductor manufacturing tool and the measurement data of the corresponding wafers.

[0012] A further improvement is that when the window moving LSTM module works, the process variables of the semiconductor manufacturing tool captured by the window moving LSTM module are fault detection classification (FDC) data, and the measurement data of the wafers are the measurement data of the first few wafers.

[0013] A further improvement is that when the window moving LSTM module works, the measurement data of the wafers obtained by the window moving LSTM module are the measurement data of the first 3 to 5 wafers.

[0014] A further improvement is that the window moving LSTM module includes a normalization module and a Principal Components Analysis (PCA) module.

[0015] The normalization module is used to normalize the FDC data.

[0016] The PCA module is used to extract features from the FDC data after standard ring processing.

[0017] A further improvement is that the PCA module extracts features from a batch of the FDC data.

[0018] A further improvement is that the PCA module is implemented by Multilinear Principal Components Analysis (MPCA).

[0019] A further improvement is that the window moving LSTM module calculates the long-term drift information of the LSTM module before update according to the measurement data of the wafers, and the long-term drift information is the residual value between the measurement data of the wafers and the predicted value of the measurement data of the LSTM module before update.

[0020] A further improvement is that the window moving LSTM module further includes a measurement data sequential input module. When the window moving LSTM module works, the measurement data sequential input module inputs the measurement data of the wafer into the LSTM module before update in sequence.

[0021] A further improvement is that the LSTM module includes a plurality of LSTM model units.

[0022] When updating the model parameters of the LSTM module, the window moving LSTM module updates the model parameters of each LSTM model unit in sequence, so as to obtain the updated LSTM module.

[0023] A further improvement is that when the LSTM module works, the data of the semiconductor manufacturing machine tool input is subjected to feature extraction by a PCA module.

[0024] A further improvement is that the semiconductor manufacturing machine tool includes a thin film deposition machine tool.

[0025] A further improvement is that the thin film deposition machine tool includes a CVD machine tool.

[0026] The CVD machine tool is used for depositing a dielectric film.

[0027] The dielectric film includes silicon oxide and silicon nitride.

[0028] The measurement data of the wafer includes the thickness measurement data of the dielectric film deposited on the wafer.

[0029] A further improvement is that the window moving LSTM module works after the semiconductor manufacturing machine tool has produced for a certain period of time, so as to make the prediction accuracy of the LSTM module stable in different time periods.

[0030] A further improvement is that the semiconductor manufacturing machine tool includes a plurality of identical process chambers, and the window moving LSTM module works when the process chambers of the semiconductor manufacturing machine tool are switched, so as to make the prediction accuracy of the LSTM module stable on different process chambers.

[0031] The present invention can automatically predict some key measurement parameters according to the production machine tool data, has a high prediction accuracy, and can improve the stability of the prediction accuracy. For example, it has a stable prediction accuracy in different process chambers and different periods, can timely detect abnormalities in the manufacturing process, save measurement benefits, and achieve batch control.

[0032] The present invention can save measurement time and measurement cost:

[0033] The model of the present invention can predict the measurement results through the machine tool Sensor data and the initial measurement data, which not only saves the actual measurement time and reduces the cycle time, but also enables real-time monitoring of the wafer quality and saves production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The present invention will be further described in detail below in conjunction with the drawings and specific embodiments:

[0035] Figure 1 is a schematic structural diagram of a virtual metrology device for a semiconductor manufacturing machine tool based on LSTM in an embodiment of the present invention;

[0036] Figure 2 is a schematic diagram of the chain structure of LSTM. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0037] As Figure 1 shown, it is a schematic structural diagram of a virtual metrology device for a semiconductor manufacturing machine tool based on LSTM in an embodiment of the present invention; the virtual metrology device for a semiconductor manufacturing machine tool based on LSTM in an embodiment of the present invention includes:

[0038] An LSTM module for outputting a predicted value of measurement data according to the input data of the semiconductor manufacturing machine tool; the predicted value of measurement data is the predicted value of the measurement data of the wafer that has completed the manufacturing process of the semiconductor manufacturing machine tool.

[0039] A window moving LSTM module for capturing the long-term drift information of the process variables and measurement data of the semiconductor manufacturing machine tool to update the model parameters of the LSTM module, thereby improving the stability of the prediction accuracy of the LSTM module.

[0040] Figure 1 In, the LSTMⅠ corresponding to the label 103 represents the LSTM module before update, and the LSTMⅡ corresponding to the label 110 represents the LSTM module after update.

[0041] In an embodiment of the present invention, the initial model parameters of the LSTM module are obtained by learning the known data of the semiconductor manufacturing machine tool and the corresponding measurement data of the wafer.

[0042] When the window moving LSTM module works, the process variables of the semiconductor manufacturing machine tool captured by the window moving LSTM module are FDC data, and the measurement data of the wafer is the measurement data of the previous several wafers.

[0043] Figure 1 In, the FDC data corresponding to the label 104 represents FDC data, and the FDC data receives the real-time sensing data from different sensors coupled to the semiconductor manufacturing machine tool.

[0044] The measurement data corresponding to the marker 101 represents the measurement data of the wafer. When the window moving LSTM module works, the measurement data of the wafer obtained by the window moving LSTM module is the measurement data of the previous 3 to 5 wafers.

[0045] The window moving LSTM module includes a normalization module 105 and a PCA module 106.

[0046] The normalization module 105 is used to normalize the FDC data. Figure 1 In, the normalization module 105 is also represented by Min-max normalization.

[0047] The PCA module 106 is used to extract features from the FDC data after standard ring processing. In the embodiment of the present invention, the PCA module 106 extracts features from a batch of the FDC data. Figure 1 In, the PCA module 106 is also represented by Batch-wise PCA.

[0048] In some preferred embodiments, the PCA module 106 is implemented by MPCA.

[0049] The PCA module 106 implements PCA analysis, that is, principal component analysis, aiming to use the idea of dimensionality reduction to convert multiple indicators into a few comprehensive indicators; among them, MPCA is multi-linear principal component analysis, which is an extension of PCA to multiple dimensions. PCA is the projection of a vector to a vector, while MPCA is the projection of a tensor to a tensor, and the projection structure is relatively simple. In addition, the operation is performed in a lower-dimensional space, so there is an advantage of low computational complexity when processing high-dimensional data.

[0050] In the embodiment of the present invention, the window moving LSTM module further includes a measurement data sequential input module 102. When the window moving LSTM module works, the measurement data sequential input module 102 inputs the measurement data of the wafer into the LSTM module before update in sequence. Figure 1 In, the measurement data sequential input module 102 is also represented by Inputsequence general.

[0051] The window moving LSTM module calculates the long-term drift information of the LSTM module before update according to the measurement data of the wafer, where the long-term drift information is the residual value between the measurement data of the wafer and the predicted value of the measurement data of the LSTM module before update. In an embodiment of the present invention, the residual value is implemented by the calculation module corresponding to label 107. Figure 1 In, the calculation module 107 of the residual value is also represented by Calculate residual.

[0052] As Figure 2 shown, it is a schematic diagram of the chain structure of LSTM. The LSTM module includes a plurality of LSTM model units 201. LSTM is a time-recurrent neural network, suitable for processing and predicting important events with very long intervals and delays in time series. Information can be added to and deleted from the cell through the gate unit (gate). Whether the information passes can be selectively determined through the gate.

[0053] From Figure 2 shown, it can be seen that the cell state of the LSTM model unit 201 is implemented through three gates, namely: the forget gate 202, the input gate 203, and the output gate 205. The forget gate 202, the input gate 203, and the output gate 205 all correspond to a neural network layer, and the activation function is implemented by the sigmoid function, i.e., the σ function, and each has its own weight and bias, and the weight and bias are model parameters that need to be obtained through learning. The neural network layer 204 is used to update the cell information and is implemented by the tanh function, and the tanh function also has corresponding weights and biases, and the weights and biases are model parameters that need to be obtained through learning.

[0054] X t-1 、X t and X t+1 represent the eigenvalue of the input data at different times, and h t-1 、h t and h t+1 represent the predicted values at different times. Figure 2 In, only the specific structure of the LSTM model unit 201 corresponding to time t is shown, and the structures of the LSTM model units 201 at times t - 1 and t + 1 are the same as that at time t.

[0055] When updating the model parameters of the LSTM module, the window moving LSTM module sequentially updates the model parameters of each LSTM model unit 201, so as to obtain the updated LSTM module 110.

[0056] When the updated LSTM module 110 works, the data of the semiconductor manufacturing machine input is subjected to feature extraction by the PCA module 112.Figure 1 Among them, the data of the input semiconductor manufacturing machine tool is as shown in Next stepdata input in Tag 111. The data of the input semiconductor manufacturing machine tool can adopt FDC data. The PCA module 112 also adopts Batch-wise PCA representation.

[0057] In an embodiment of the present invention, the semiconductor manufacturing machine tool includes a thin film deposition machine tool. The thin film deposition machine tool includes a CVD machine tool.

[0058] The CVD machine tool is used for depositing a dielectric film.

[0059] The dielectric film includes silicon oxide and silicon nitride.

[0060] The measurement data of the wafer includes the thickness measurement data of the dielectric film deposited on the wafer.

[0061] The window moving LSTM module works after the semiconductor manufacturing machine tool has produced for a certain period of time, so that the prediction accuracy of the LSTM module is stable in different time periods.

[0062] The semiconductor manufacturing machine tool includes a plurality of identical process chambers. The window moving LSTM module works when the process chambers of the semiconductor manufacturing machine tool are switched, so that the prediction accuracy of the LSTM module is stable on different process chambers.

[0063] The model prediction accuracy of the embodiment of the present invention is high and meets the process requirements:

[0064] Taking the prediction of the silicon nitride thin film deposited by the CVD machine tool as an example, the mean square error predicted in the modeling stage of the embodiment of the present invention is within 6 Å, and the mean square error in the test stage is within 9 Å. The process error requirement of the recipe for the silicon nitride thin film deposition is within 20 Å, and the measurement error between the measurement machine tools is about 5 Å. Therefore, the embodiment of the present invention can meet the process requirements.

[0065] When the same model of the embodiment of the present invention is applied to predict data with a long time span, the prediction effect still remains:

[0066] The embodiment of the present invention uses the data of the first four months to establish a model and predicts the data of the next four months, and the prediction effect still remains. That is, the initial model parameters of the LSTM module of the embodiment of the present invention can be learned and trained through the data of a long time ago, and at the same time, it will not have too much impact on the prediction accuracy of the existing production data.

[0067] The embodiment of the present invention can save measurement time and measurement cost:

[0068] The model of the embodiment of the present invention can predict the measurement result through the machine platform Sensor data and the initial measurement data, which not only saves the actual measurement time and reduces the Cycle Time, but also can realize the real-time monitoring of the wafer quality and save the production efficiency.

[0069] The present invention has been described in detail through specific embodiments, but these do not constitute a limitation to the present invention. Without departing from the principle of the present invention, those skilled in the art can also make many deformations and improvements, which should also be regarded as the protection scope of the present invention.

Claims

1. An LSTM-based virtual metrology device for semiconductor manufacturing machines, characterized in that, it includes: an LSTM module for outputting a predicted value of metrology data according to the input data of the semiconductor manufacturing machine; the predicted value of metrology data is the predicted value of the metrology data of the wafer that has completed the manufacturing process of the semiconductor manufacturing machine; a window moving LSTM module for capturing the long-term drift information of the process variables and metrology data of the semiconductor manufacturing machine to update the model parameters of the LSTM module, thereby improving the stability of the prediction accuracy of the LSTM module; when the window moving LSTM module works, the process variables of the semiconductor manufacturing machine captured by the window moving LSTM module are FDC data, and the metrology data of the wafer is the metrology data of the first 3 to 5 wafers; the window moving LSTM module includes a normalization module and a PCA module; the normalization module is used to perform normalization processing on the FDC data; the PCA module is used to extract features from the FDC data after normalization processing; the window moving LSTM module calculates the long-term drift information of the LSTM module before update according to the metrology data of the wafer, and the long-term drift information is the residual value between the metrology data of the wafer and the predicted value of the metrology data of the LSTM module before update; the window moving LSTM module further includes a metrology data sequential input module. When the window moving LSTM module works, the metrology data sequential input module inputs the metrology data of the wafer into the LSTM module before update in sequence.

2. The LSTM-based virtual metrology device for semiconductor manufacturing machines according to claim 1, characterized in that: the initial model parameters of the LSTM module are obtained by learning the known data of the semiconductor manufacturing machine and the corresponding metrology data of the wafer.

3. The LSTM-based virtual metrology device for semiconductor manufacturing machines according to claim 1, characterized in that: the PCA module extracts features from a batch of the FDC data.

4. The LSTM-based virtual metrology device for semiconductor manufacturing machines according to claim 3, characterized in that: the PCA module is implemented by MPCA.

5. The LSTM-based virtual metrology device for semiconductor manufacturing machines according to claim 1, characterized in that: the LSTM module includes a plurality of LSTM model units; when updating the model parameters of the LSTM module, the window moving LSTM module sequentially updates the model parameters of each LSTM model unit to obtain the updated LSTM module.

6. The LSTM-based virtual metrology device for semiconductor manufacturing machines according to claim 1 or 5, characterized in that: when the updated LSTM module works, the input data of the semiconductor manufacturing machine is subjected to feature extraction by the PCA module.

7. The LSTM-based virtual metrology device for semiconductor manufacturing machines according to claim 1, characterized in that: The semiconductor manufacturing tool includes a thin film deposition tool.

8. The virtual metrology device for a semiconductor manufacturing tool based on LSTM according to claim 7, wherein: the thin film deposition tool includes a CVD tool; the CVD tool is used for depositing a dielectric film; the dielectric film includes silicon oxide and silicon nitride; the metrology data of the wafer includes the thickness metrology data of the dielectric film deposited on the wafer.

9. The virtual metrology device for a semiconductor manufacturing tool based on LSTM according to claim 1, wherein: the window moving LSTM module operates after the semiconductor manufacturing tool has been producing for a certain period of time, so that the prediction accuracy of the LSTM module is stable in different time periods.

10. The virtual metrology device for a semiconductor manufacturing tool based on LSTM according to claim 1, wherein: the semiconductor manufacturing tool includes a plurality of identical process chambers, and the window moving LSTM module operates when the process chambers of the semiconductor manufacturing tool are switched, so that the prediction accuracy of the LSTM module is stable on different process chambers.

Images