A film thickness prediction method and system based on self-supervision and representation learning

Abstract

The present invention discloses a method and system for predicting film thickness based on self-supervision and representation learning, which belongs to the field of machine learning and artificial intelligence technology. A labeled first training set and an unlabeled second training set are obtained, the encoder is pre-trained with the second training set, and then fine-tuned with the first training set, while training the dynamic feature selector and the regression head. The dynamic feature selector reduces the dimension of the first training set data, and then uses the encoder to initialize the global prototype projection space. The encoder generates an embedded representation of the reduced-dimensional data, the feature representation head calculates the projection coordinates, and the projection loss, regression loss, diversity loss and global prototype orthogonal loss are combined to update the model; finally, the actual data is used as input to predict the film thickness through the dynamic feature selector, encoder and regression head. The present invention does not need to rely on a large amount of labeled data, solves the feature selection problem, representation aliasing and localization problem of highly complex semiconductor manufacturing data, and has high prediction accuracy and stability.

Description

Technical Field

[0001] The present invention belongs to the field of machine learning and artificial intelligence technology, and in particular relates to a film thickness prediction method and system based on self-supervision and representation learning. Background Art

[0002] In the field of semiconductor manufacturing, precise control of the production process is crucial to ensuring high-quality products and operational efficiency. Chemical Vapor Deposition (CVD) is one of the key processes for thin film preparation, and its process quality directly affects the reliability of semiconductor devices. However, given the diversity and complexity of process variables and the potential performance drift of the equipment itself, the uniformity of film thickness remains a very challenging problem. Traditional virtual metrology (VM) technology mainly relies on prediction methods based on physical models or machine learning, while existing deep learning prediction methods have the following three problems:

[0003] (1) Dependence on a large amount of labeled data: The accuracy of film thickness prediction by current deep learning methods is highly dependent on a large amount of labeled data. However, since the measurement process is time-consuming and costly, it is extremely difficult to obtain labeled data.

[0004] (2) Feature selection problem: Feature selection methods are mainly divided into three categories: filtering, wrapping, and embedding. The filtering method is computationally simple, but does not consider the interaction between features and is difficult to apply to highly complex semiconductor manufacturing data. The wrapping method selects features by training the model through multiple iterations, which has a high computational cost. It is suitable for traditional machine learning methods but not for deep learning. Existing deep learning feature screening methods lack in-depth research on embedding methods, that is, the lack of dynamic feature screening mechanisms has led to many deep learning models directly using raw sensor data for prediction and failing to effectively screen key features. This may cause redundant information to interfere with the prediction results and reduce the generalization ability of the model.

[0005] (3) Representation aliasing and localization issues: Feature representations learned by existing deep learning methods often suffer from aliasing and lack clear separability, which affects the clarity and accuracy of decision making. In addition, the feature representation of each sample is usually independent, ignoring the correlation of the global data structure, resulting in unstable prediction results. Summary of the Invention

[0006] In response to the above problems, the present invention proposes a thin film thickness prediction method and system based on self-supervision and representation learning. It uses unlabeled data to pre-train the encoder, reduces dependence on labeled data, enhances data features and improves the generalization ability of the model. It also improves prediction accuracy by constructing a dynamic selector through embedded feature selection. In addition, it optimizes the representation space by constructing a global prototype projection space, reduces representation aliasing and localization problems, and makes the prediction results more stable. With the help of the optimal transmission mechanism and orthogonality constraints, it improves the prediction reliability and the adaptability of the model to different data distributions.

[0007] The technical solution adopted in the present invention is as follows:

[0008] In a first aspect, the present invention proposes a film thickness prediction method based on self-supervision and representation learning, comprising the following steps:

[0009] (1) Collecting wafer manufacturing raw process data and film thickness measurement data, pre-processing the data pairs that successfully match the raw process data and measurement data as the first training set, and pre-processing the raw process data that do not successfully match as the second training set;

[0010] (2) Pre-train an encoder using the second training set, then fine-tune the encoder using the first training set and simultaneously train a dynamic feature selector and a regression head;

[0011] (3) Using a dynamic feature selector to reduce the dimension of the process data in the first training set, and using the fine-tuned encoder to initialize a global prototype projection space for the process data after dimension reduction;

[0012] (4) The newly updated encoder is used to generate an embedded representation of the process data after dimensionality reduction. The projection coordinates of the embedded representation in the global prototype projection space are calculated through the feature representation head. The projection loss is calculated based on the projection distribution, and the regression loss is calculated based on the regression head. At the same time, the diversity loss and the global prototype orthogonal loss are introduced. The encoder, regression head, and feature representation head are iteratively updated with the combined total loss, and the global prototype projection space is updated synchronously.

[0013] (5) Taking the actual wafer manufacturing process data as input, the film thickness is predicted using a dynamic feature selector, an encoder, and a regression head.

[0014] Furthermore, the preprocessing refers to preliminary screening of features in the process data, and the screened features include deposition time, buffer pressure, wafer processing temperature, lateral and top silane gas reaction time, and top, bias and lateral RF reflection power.

[0015] Furthermore, when the second training set is used to pre-train an encoder, a K-Means clustering method is used to generate pseudo measurement data of the second training set as pseudo labels for training, and the process data in the second training set is perturbed;

[0016] Furthermore, reusing the first training set to fine-tune the encoder and synchronously training a dynamic feature selector and a regression head includes:

[0017] Using a dynamic feature selector to filter the features of the process data of the first training set to obtain the process data after dimensionality reduction;

[0018] Use pre-trained feature encoders to generate embedding representations of the process data after dimensionality reduction;

[0019] Predicting film thickness based on the embedded representation using a regression head;

[0020] The loss is calculated based on the predicted film thickness and the actual measurement data, and the encoder, dynamic feature selector and regression head parameters are updated.

[0021] Furthermore, the dynamic feature selector is specifically:

[0022] Introducing random gating variables , gate value Defined as:

[0023] ;

[0024] in, is the jth trainable parameter in the dynamic feature selector, j = 1, 2, ... d; d represents the number of features of the samples in the training set; is an independent sampling parameter of the jth trainable parameter, which obeys the Gaussian distribution; by calculating the process data and the random gate variable The Hadamard product is used to realize feature screening, and the regularization term is introduced into the training process of the dynamic feature selector.

[0025] Furthermore, the initialization process of the global prototype projection space is:

[0026] Generate an embedding representation of the reduced-dimensionality process data using the fine-tuned encoder.

[0027] The embedded representations are clustered, and each cluster center is regarded as a prototype, resulting in a global prototype projection space containing several prototypes.

[0028] Furthermore, the projection coordinates of the embedding representation in the global prototype projection space calculated by the feature representation head are expressed as:

[0029] ;

[0030] Embedded representation The projected coordinates of The following relationship is satisfied with the global prototype projection space:

[0031] ;

[0032] in, represents the number of prototypes in the global prototype projection space, represents the k-th prototype in the global prototype projection space, Indicates the feature indicates the head.

[0033] Furthermore, the diversity loss calculation process is as follows:

[0034] The range of the measured data in the first training set is divided into several subintervals. The process data in the same subinterval are regarded as a pair of positive samples, and the process data in different subintervals are regarded as a pair of negative samples. The contrastive learning loss of the projection coordinates is calculated based on the positive and negative samples.

[0035] Furthermore, the projection loss refers to the optimal transmission distance loss between the distribution of the projection coordinates and the distribution of the embedded representation, and the global prototype orthogonality loss refers to the global prototype projection space satisfying the orthogonality constraint.

[0036] In a second aspect, the present invention proposes a film thickness prediction system based on self-supervision and representation learning, which is used to implement the above-mentioned film thickness prediction method based on self-supervision and representation learning.

[0037] The beneficial effects of the present invention are:

[0038] The present invention first uses unlabeled wafer manufacturing process data to pre-train the encoder, which can not only reduce the dependence on a large amount of labeled data, but also enhance the data features, thereby improving the model's generalization ability; an embedded feature selection method is used to construct a dynamic feature selector, so that the model can automatically screen the most representative wafer manufacturing process features, reduce redundant information interference, and effectively improve prediction accuracy; in terms of representation space optimization, by constructing a global prototype projection space, the separability of the embedded representation of wafer manufacturing process data is improved, the representation aliasing problem is reduced, and the film thickness prediction results are more stable; with the help of the optimal transmission mechanism, the optimal transmission loss optimization is achieved, which enables the model to more effectively learn the global structural information of the data and improve the reliability of the prediction; the orthogonality constraint method is used to implement orthogonality constraints to improve robustness, make the global prototype more independent, and enhance the model's adaptability to different data distributions. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Flowchart of the film thickness prediction method based on self-supervision and representation learning;

[0040] Figure 2 The framework diagram of the film thickness prediction method based on self-supervision and representation learning;

[0041] Figure 3 Schematic diagram of the self-supervised pre-training stage;

[0042] Figure 4 Schematic diagram of the fine-tuning training stage;

[0043] Figure 5 Schematic diagram of the characterization learning stage. DETAILED DESCRIPTION

[0044] The following description is provided to disclose the present invention so that those skilled in the art can implement the present invention.

[0045] The accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Some of the blocks shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0046] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all steps. For example, some steps may be decomposed, while some steps may be combined or partially combined, so the actual execution order may change according to actual circumstances.

[0047] The film thickness prediction method based on self-supervision and representation learning proposed in this paper is named SGT_SRL (Self-Supervised and Representation-Aware Transformer). The overall framework structure is as follows: Figure 2 As shown in FIG, the self-supervised pre-training, fine-tuning training and representation learning stages are combined, which can effectively use CVD process data to predict film thickness. The training process of the three stages will be detailed below.

[0048] like Figure 1 As shown, the implementation process of the present invention mainly includes the following steps:

[0049] S1, obtaining a first training set and a second training set for model training.

[0050] Collect real-time data from the fault detection and classification system (FDC system) during wafer manufacturing, including process data and measurement data of key semiconductor manufacturing steps. The process data is regarded as process variables and the measurement data is regarded as target variables.

[0051] In the specific implementation, the lateral / top silane of the CVD process was collected ( ) Parameters such as gas flow rate, top / bias / side RF reflected power, deposition time, and front pressure are used as process data, and the final film thickness is used as measurement data.

[0052] The data pairs of original process data and measurement data that are successfully matched are pre-processed as the first training set, and the original process data that are not successfully matched are pre-processed as the second training set. The two data sets are used for subsequent SGT_SRL training. The SGT_SRL training process includes three stages: self-supervised pre-training stage, fine-tuning training stage, and representation learning stage. The model of the present invention is defined as , whose parameters are and ,in By dynamic feature selector , encoder And the return header composition.

[0053] S2, self-supervised pre-training stage: pre-train an encoder using the second training set .

[0054] like Figure 3 As shown, this stage requires creating a pre-training task from the unlabeled second training set. In actual situations, there is a high correlation between certain input process data features and the original measurement data. For example, the original task is to use features TOP_MEAN_1, SID_MEAN_2, and SID_MEAN_3 to predict the film thickness BTH. The deposition time TIME_1 is highly correlated with the film thickness BTH, so the features TOP_MEAN_1, SID_MEAN_2, and SID_MEAN_3 can be used to predict the deposition time TIME_1 to create a pre-training task. In order to improve the diversity and rationality of pseudo-labels, this embodiment will first select the top 10 most correlated features for clustering to generate pseudo-labels.

[0055] First, define a binary mask matrix:

[0056]

[0057] Among them, d is the feature dimension of the input data, which is 81. , Indicates that the feature is selected. Indicates that the feature is not selected. The features selected by the mask form the new input:

[0058]

[0059] in, represents element-wise multiplication, Indicates removing elements with mask value 0. Based on the selected features, run K-Means clustering to generate task labels:

[0060]

[0061] in, Indicates the i-th original process data that was not successfully matched, represents the binary mask matrix used to preprocess the raw process data, Represents the features of the preprocessed process data, with a feature dimension of w; represents the cluster center matrix, represents the pseudo label of the i-th process data in the second training set, which is in the form of a one-hot encoding with dimension ks, and , is a vector of all 1s, represents the number of samples in the second training set, Represents the square of the L2 norm. is the number of cluster centers, set to 100.

[0062] In order to prevent the model from memorizing data directly, random variables are introduced:

[0063]

[0064]

[0065] in, It follows a Bernoulli distribution with a parameter of 0.5, where the parameter 0.5 refers to the probability that the random variable takes the value 1; the superscript Indicates transpose.

[0066] for The position with value 1 is replaced by the mean of the feature column to obtain the perturbed input Finally, the pre-training task dataset is constructed and passed into the model for pre-training:

[0067]

[0068] The process of pre-training the model is: using the perturbed process data As input to the encoder, a classification head predicts pseudo labels for the process data. The encoder and classification head are updated based on the predicted results and the actual pseudo labels. Here, the encoder can use a Transformer network, and the classification head can be implemented using one or more linear layers. The classification head is only used during training at this stage.

[0069] S3, fine-tuning training phase: fine-tune the encoder using the first training set and simultaneously train a dynamic feature selector and a regression head.

[0070] This stage uses the first training set for training.

[0071] like Figure 4 As shown, the encoder The encoder pre-trained in the previous stage is used to initialize the parameters. In order to select features, an embedded method is used. Specifically, a random gating variable is introduced. , to achieve feature screening, for each feature Perform independent sampling:

[0072]

[0073] in, is a fixed noise intensity, the gate value The definition is as follows:

[0074]

[0075] in, For feature selector The trainable parameters of When , the features are fully preserved, When , the feature is completely suppressed. After the input process data is filtered by features, it is passed to the encoder and regression head to calculate the regression loss:

[0076]

[0077] In order to enhance feature sparsity, a regularization term is introduced:

[0078]

[0079] in, is the cumulative distribution function (CDF) of the standard normal distribution, Used to adjust the weight of the regularization term in the overall loss. This regularization term encourages non-important features to , thereby closing the corresponding gate. Based on the results of the loss function, update the model parameters and :

[0080]

[0081]

[0082] in, and Corresponding to the encoder and regression header parameter With feature selector of The learning rate is set, and the above steps are repeated iteratively to complete the fine-tuning step.

[0083] S4, representation learning stage.

[0084] like Figure 5 As shown in Figure 2, this stage includes the global prototype projection space initialization process, as well as the process of iteratively updating the encoder, regression head, and feature representation head.

[0085] First, in this stage, the prototype-based projection space (P-Space) is constructed to optimize the data representation method, improve the generalization ability of the model, and reduce the aliasing problem between different data features. and , and get the embedded representation ,here Indicates the number of training samples in a batch, where training samples refer to the samples in the first training set. Perform K-Means clustering to obtain the cluster center matrix , and use it to initialize the global prototype:

[0086]

[0087] in, is the number of clusters, and the cluster center matrix of each cluster represents a vector of length s, which is recorded as a prototype; Represents the kth prototype. Define the prototype projection space as the global prototype As the projection space of basis vectors, the embedding representation of each sample is Formalized as , .

[0088] In this embodiment, the embedding representation of the i-th sample is Representing the head through features Calculate the projection coordinates in the prototype projection space , feature representation head This can be achieved using one or more linear layers.

[0089] The projected distribution of the embedded representation and projection coordinates are expressed as:

[0090]

[0091]

[0092] in, express The Dirac function, express The Dirac function, and are the projected distributions of the embedded representation and projection coordinates, respectively.

[0093] To make the prototype Capture global data structure information and use the optimal transmission (OT) distance to calculate the projection loss:

[0094]

[0095] Where N is the number of samples.

[0096] Considering that direct optimization may lead to coordinate collapse, constraints are introduced to distribute the projected coordinates to different areas. In this embodiment, the range of the measurement data is divided into t sub-intervals, where , is the batch size. Samples within the same partition are considered positive pairs, and cross-region samples are considered negative pairs. The optimization objective is:

[0097]

[0098] in, is the indicator function, when Positive sample pair When is 1.

[0099] In addition, the global prototype projection space is composed of global prototypes. In order to make the prototypes independent, the orthogonality constraint must be satisfied:

[0100]

[0101] in, , the first item Force M to become sparse, and the second term ensures . Will task loss and Combined and iterative updates Parameters, you can get the final model for prediction .

[0102] It should be noted that the above only gives The calculation formula of Refers to the conventional regression loss, that is, the loss between the predicted film thickness and the actual measured data.

[0103] S5. Taking the actual wafer manufacturing process data as input, the film thickness is predicted using a dynamic feature selector, an encoder, and a regression head.

[0104] Here, the input process data is first filtered through a dynamic feature selector to obtain the process data after dimensionality reduction. Then, a feature encoder is used to generate an embedded representation of the process data after dimensionality reduction. Finally, a regression head is used to predict the film thickness based on the embedded representation.

[0105] In order to quantitatively evaluate the performance of the SGT_SRL model, the following experimental verification is performed.

[0106] Baseline and evaluation metric selection: Representative methods from machine learning, deep learning, and self-supervised learning are selected for comparative evaluation, using the root mean square error (RMSE) and the coefficient of determination (R 2 ) as the evaluation indicator.

[0107] Deep learning methods include FT-Transformer, GNN, NODE, MDN, AutoInt, GANDALF, MLP, ResNet, TabDPT, STG, and TabTransformer. They have achieved good performance on public datasets of tabular data, and achieve excellent nonlinear modeling capabilities by incorporating self-attention mechanisms and stacking multi-layer networks.

[0108] Machine learning methods include LightGBM, XGBoost, Random Forest, CatBoost, Extra-Trees, GPR, PLS, SVR, Elastic Net, and Linear Regression (LR). These methods are often used in traditional virtual measurement applications because they require little training time and offer relatively stable prediction performance.

[0109] Self-supervised learning methods include TabNet, VIME, SAINT, and SCARF. They enhance the generalization ability of the model's feature representation by generating pseudo labels on unlabeled data to generate pre-training tasks.

[0110] Data Collection and Preprocessing: 12-inch wafer production data was collected, including 81 characteristic variables such as reaction gas flow, RF power, temperature, and chamber pressure. The data was normalized and processed for missing values. The data was divided into four data sets (CVD-1, CVD-2, CVD-3, and CVD-4) based on the different tools used. These data sets were then divided into training, validation, and test sets, and fed into the comparison model for evaluation. The experimental results are shown in Tables 1 and 2.

[0111] Table 1:

[0112]

[0113] Table 2:

[0114]

[0115] Comparing Table 1 and Table 2, we can see that SGT_SRL achieved RMSE values ​​of 18.796, 26.908, 19.710, and 21.730, and R 2 The values ​​are 0.769, 0.590, 0.649 and 0.640. Except for the CVD2 dataset, where it performs slightly worse, SGT_SRL outperforms all the compared methods in all other cases.

[0116] Ablation experiments were conducted on the present invention, and the results are shown in Tables 3 and 4.

[0117] Table 3:

[0118]

[0119] Table 4:

[0120]

[0121] Tables 3 and 4 show that the complete model (FTT_SG(SL, RL)) using SL+RL+feature selection performs best on all datasets. Representation learning (RL) is a key factor in improving model performance, while self-supervised learning (SL) contributes significantly on some datasets. Feature selection can improve model performance, but it requires a combination of RL and SL to maximize the effect.

[0122] The present invention also provides a film thickness prediction system based on self-supervision and representation learning, which is used to implement the above-mentioned embodiments. The terms "module," "unit," etc. used below may refer to a combination of software and / or hardware that implements a predetermined function. Although the systems described in the following embodiments are preferably implemented in software, implementation using hardware, or a combination of software and hardware, is also possible.

[0123] This embodiment provides a film thickness prediction system based on self-supervision and representation learning, including:

[0124] A training data preprocessing module is used to collect wafer manufacturing raw process data and film thickness measurement data, preprocess the data pairs that successfully match the raw process data and the measurement data as the first training set, and preprocess the raw process data that do not successfully match as the second training set;

[0125] a pre-training module for pre-training an encoder using a second training set;

[0126] A fine-tuning module, configured to fine-tune the encoder obtained by the pre-training module using the first training set and simultaneously train a dynamic feature selector and a regression head;

[0127] A global prototype projection space initialization module is used to reduce the dimension of the process data in the first training set using a dynamic feature selector, and initialize a global prototype projection space using the fine-tuned encoder for the process data after the dimension reduction process;

[0128] The representation learning module is used to generate an embedded representation of the process data after dimensionality reduction using the newly updated encoder. The feature representation head calculates the projection coordinates of the embedded representation in the global prototype projection space. The projection loss is calculated based on the optimal transmission mechanism of the projection distribution. The regression loss is calculated based on the regression head. Diversity loss and global prototype orthogonal loss are introduced at the same time. The encoder, regression head, and feature representation head are iteratively updated with the combined total loss, and the global prototype projection space is also updated synchronously.

[0129] The film thickness prediction module is used to predict the film thickness using the actual wafer manufacturing process data as input and using a dynamic feature selector, an encoder and a regression head.

[0130] As for the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can be referred to the partial description of the method embodiment, and the implementation methods of the remaining modules will not be repeated here. The system embodiment described above is only illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present invention. Ordinary technicians in this field can understand and implement it without paying any creative work.

[0131] Embodiments of the system of the present invention can be applied to any device with data processing capabilities, such as a computer or other device. System embodiments can be implemented through software, hardware, or a combination of software and hardware. For example, a software implementation, as a logical device, is implemented by a processor of any device with data processing capabilities, reading corresponding computer program instructions from non-volatile memory into internal memory and executing them.

[0132] The above-described embodiments merely illustrate several implementations of the present invention, and while the descriptions are relatively specific and detailed, they are not to be construed as limiting the scope of the present invention. Persons skilled in the art will readily appreciate that variations and modifications may be made without departing from the spirit of the present invention, and all such variations and modifications fall within the scope of protection of the present invention.

Claims

1. A film thickness prediction method based on self-supervision and representation learning, characterized in that: The following steps are involved: (1) Collecting wafer manufacturing raw process data and film thickness measurement data, pre-processing the data pairs that successfully match the raw process data and measurement data as the first training set, and pre-processing the raw process data that do not successfully match as the second training set; (2) Pre-train an encoder using the second training set, then fine-tune the encoder using the first training set and simultaneously train a dynamic feature selector and a regression head; (3) Using a dynamic feature selector to reduce the dimension of the process data in the first training set, and using the fine-tuned encoder to initialize a global prototype projection space for the process data after dimension reduction; (4) The newly updated encoder is used to generate an embedded representation of the process data after dimensionality reduction. The projection coordinates of the embedded representation in the global prototype projection space are calculated through the feature representation head. The projection loss is calculated based on the optimal transmission mechanism of the projection distribution. The regression loss is calculated based on the regression head. At the same time, the diversity loss and the global prototype orthogonal loss are introduced. The encoder, regression head, and feature representation head are iteratively updated with the combined total loss, and the global prototype projection space is updated synchronously. (5) Taking the actual wafer manufacturing process data as input, the film thickness is predicted using a dynamic feature selector, an encoder, and a regression head.

2. The thin film thickness prediction method based on self-supervision and representation learning according to claim 1, characterized in that: The pre-processing is to perform preliminary screening of features in the process data, wherein the screened features include deposition time, buffer pressure, wafer processing temperature, lateral and top silane gas reaction time, and top, bias and lateral RF reflection power.

3. The thin film thickness prediction method based on self-supervision and representation learning according to claim 1, characterized in that: When pre-training an encoder using the second training set, a K-Means clustering method is used to generate pseudo measurement data of the second training set as pseudo labels for training, while perturbing the process data in the second training set; The optimization goal of generating pseudo labels is: ； in, Indicates the i-th original process data that was not successfully matched, represents the binary mask matrix used to preprocess the raw process data, Represents the features of the preprocessed process data, with a feature dimension of w; represents the cluster center matrix, represents the pseudo label of the i-th process data in the second training set, which is in the form of a one-hot encoding with dimension ks, and , is a vector of all 1s, represents the number of samples in the second training set, Represents the square of the L2 norm.

4. The thin film thickness prediction method based on self-supervision and representation learning according to claim 1, characterized in that: The method of reusing the first training set to fine-tune the encoder and synchronously training a dynamic feature selector and a regression head includes: Using a dynamic feature selector to filter the features of the process data of the first training set to obtain the process data after dimensionality reduction; Use pre-trained feature encoders to generate embedding representations of the process data after dimensionality reduction; Predicting film thickness based on the embedded representation using a regression head; The loss is calculated based on the predicted film thickness and the actual measurement data, and the encoder, dynamic feature selector and regression head parameters are updated.

5. The thin film thickness prediction method based on self-supervision and representation learning according to claim 4, characterized in that: The dynamic feature selector is specifically: Introducing random gating variables , gate value Defined as: ； in, is the jth trainable parameter in the dynamic feature selector, j = 1, 2, ... d; d represents the number of features of the samples in the training set; is an independent sampling parameter of the jth trainable parameter, which obeys the Gaussian distribution; by calculating the process data and the random gate variable The Hadamard product is used to realize feature screening, and the regularization term is introduced into the training process of the dynamic feature selector.

6. The thin film thickness prediction method based on self-supervision and representation learning according to claim 1, characterized in that: The initialization process of the global prototype projection space is: Generate an embedding representation of the reduced-dimensionality process data using the fine-tuned encoder. The embedded representations are clustered, and each cluster center is regarded as a prototype, resulting in a global prototype projection space containing several prototypes.

7. The thin film thickness prediction method based on self-supervision and representation learning according to claim 6, characterized in that: The projection coordinates of the embedding representation in the global prototype projection space calculated by the feature representation head are expressed as: ； Embedded representation The projected coordinates of The following relationship is satisfied with the global prototype projection space: ； in, represents the number of prototypes in the global prototype projection space, represents the k-th prototype in the global prototype projection space, Indicates the feature indicates the head.

8. The thin film thickness prediction method based on self-supervision and representation learning according to claim 1 or 7, characterized in that: The diversity loss calculation process is as follows: The range of the measured data in the first training set is divided into several subintervals. The process data in the same subinterval are regarded as a pair of positive samples, and the process data in different subintervals are regarded as a pair of negative samples. The contrastive learning loss of the projection coordinates is calculated based on the positive and negative samples.

9. The thin film thickness prediction method based on self-supervision and representation learning according to claim 1 or 7, characterized in that: The projection loss refers to the optimal transmission distance loss between the distribution of the projection coordinates and the distribution of the embedded representation, and the global prototype orthogonality loss refers to the global prototype projection space satisfying the orthogonality constraint.

10. A film thickness prediction system based on self-supervision and representation learning, used to implement the film thickness prediction method based on self-supervision and representation learning according to claim 1, characterized in that: The system comprises: A training data preprocessing module is used to collect wafer manufacturing raw process data and film thickness measurement data, preprocess the data pairs that successfully match the raw process data and the measurement data as the first training set, and preprocess the raw process data that do not successfully match as the second training set; a pre-training module for pre-training an encoder using a second training set; A fine-tuning module, configured to fine-tune the encoder obtained by the pre-training module using the first training set and simultaneously train a dynamic feature selector and a regression head; A global prototype projection space initialization module is used to reduce the dimension of the process data in the first training set using a dynamic feature selector, and initialize a global prototype projection space using the fine-tuned encoder for the process data after the dimension reduction process; The representation learning module is used to generate an embedded representation of the process data after dimensionality reduction using the newly updated encoder. The feature representation head calculates the projection coordinates of the embedded representation in the global prototype projection space. The projection loss is calculated based on the optimal transmission mechanism of the projection distribution. The regression loss is calculated based on the regression head. Diversity loss and global prototype orthogonal loss are introduced at the same time. The encoder, regression head, and feature representation head are iteratively updated with the combined total loss, and the global prototype projection space is also updated synchronously. The film thickness prediction module is used to predict the film thickness using the actual wafer manufacturing process data as input and using a dynamic feature selector, an encoder and a regression head.

Images