Method and system for predicting step-level cycle time for semiconductor wafer manufacturing

Abstract

The application provides a kind of process level cycle time prediction method and system for semiconductor wafer manufacturing, it is related to semiconductor wafer manufacturing technical field, including obtaining the multi-source data of semiconductor wafer manufacturing entire process;Based on random forest model analysis multi-source data relative to process level cycle time The importance and filter out key feature set;Through cycle time prediction model processing key feature set, obtain process level cycle time prediction value;Cycle time prediction model includes regression integrated sub-model and deep embedding clustering sub-model, deep embedding clustering sub-model handles key features and obtains embedding representation and generates clustering soft assignment probability based on embedding representation, regression integrated sub-model processes embedding representation and obtains multiple local cycle time prediction values, and regression integrated sub-model converts clustering soft assignment probability into clustering weight by built-in dynamic modulation mechanism.The application realizes the accurate prediction of process level cycle time in semiconductor wafer manufacturing.

Description

Technical Field

[0001] This invention relates to the field of semiconductor wafer manufacturing technology, and specifically to a method and system for predicting cycle time at the step level in semiconductor wafer manufacturing. Background Technology

[0002] As competition in the semiconductor manufacturing industry intensifies, ensuring on-time delivery of customer orders is becoming increasingly crucial. Accurate cycle time (CT) forecasts not only provide key information such as order progress, daily output, or monthly output to assist in making customer engagement decisions, but also help production control personnel identify production schedule deviations and adjust production plans accordingly.

[0003] Existing research typically employs production simulation methods (such as discrete event simulation) to construct a digital model of the wafer manufacturing system, simulate the flow of wafers between various processes and the interaction behavior of equipment, and predict CT values ​​by adjusting input variables (such as batch arrival rate, equipment parameters, scheduling rules, etc.).

[0004] However, the simulation models of production simulation methods require high computational costs and are difficult to capture dynamic changes in the production site in real time (such as sudden equipment failures, emergency order insertions, and scheduling rule adjustments), which cannot provide timely support for real-time production decisions. In addition, existing simulations mostly focus on the overall cycle prediction of wafer batches, and the modeling of fine-grained parameters such as dwell time and queuing time of individual steps is coarse, which is difficult to meet the needs of precise control at the step level. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a step-level cycle time prediction method and system for semiconductor wafer manufacturing, which solves the problems of existing methods lacking real-time performance in CT prediction and failing to meet the requirements of precise step-level control.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a step-level cycle time prediction method for semiconductor wafer manufacturing, comprising: Acquire multi-source data throughout the entire semiconductor wafer manufacturing process; The importance of the multi-source data to the step-level cycle time was analyzed based on the random forest model, and key features were selected according to the degree of importance. The key feature set is processed by a cycle time prediction model to obtain step-level cycle time prediction values. The cycle time prediction model includes a regression ensemble sub-model and a deep embedding clustering sub-model. The deep embedding clustering sub-model processes the key features to obtain an embedding representation and generates a soft clustering probability based on the embedding representation. The regression ensemble sub-model processes the embedding representation to obtain multiple local cycle time prediction values. The regression ensemble sub-model converts the soft clustering probability into clustering weights through a built-in dynamic modulation mechanism. The clustering weights and multiple local cycle time prediction values ​​are weighted and fused to generate the step-level cycle time prediction values.

[0007] Preferably, the regression ensemble sub-model converts the clustering soft assignment probabilities into clustering weights through a built-in dynamic modulation mechanism, including: The correction vector is calculated using the multilayer perceptron in the regression ensemble sub-model based on the cluster soft assignment probability. Add the clustering soft assignment probability and the correction vector; The clustering weights are obtained by renormalizing the clustering soft assignment probabilities and the sum of the correction vectors using a vector-level normalized activation function.

[0008] Preferably, the deep embedding clustering sub-model includes a pre-training phase, which includes: Step S210: Obtain the key feature set; preset the balance parameters and number of clusters for the deep embedding clustering sub-model. Initialize the encoder weights and decoder weights of the stacked autoencoders in the deep embedding clustering sub-model; Step S220: Process the key feature set through the stacked autoencoder to obtain an embedded representation; Step S230: Cluster the embedding representation using a clustering algorithm to initialize... Cluster centers ; Step S240: Calculate the total loss based on the balance parameters, the current embedding representation, and the cluster centers; Step S250: Based on the total loss, update the cluster centers, encoder weights and decoder weights in reverse using the gradient descent algorithm, and obtain a new embedding representation based on the updated stacked autoencoder; Step S260: Repeat steps S240 to S250 until the maximum number of iterations is reached to obtain the parameters of the pre-trained deep embedding clustering sub-model.

[0009] Preferably, after pre-training the deep embedding clustering sub-model, the following is included: Obtain the embedding representation and cluster soft assignment probability output by the deep embedding clustering sub-model; Remove the decoder part of the deep embedding clustering sub-model and construct A regression subnetwork is created and connected to the encoder; Initialize the weights of the regression sub-networks and the parameters of the multilayer perceptron in the regression ensemble sub-model; The periodic time prediction is obtained by processing the embedded representation and soft-assigned probability through a regression ensemble sub-model, and the prediction loss is calculated based on the periodic time prediction and the actual periodic time. Based on the predicted loss, the cluster centers, autoencoder weights, regressor subnetwork weights, and multilayer perceptron parameters are updated in reverse using the gradient descent algorithm to obtain a fine-tuned periodic time prediction model.

[0010] Preferably, the total loss includes reconstruction loss and clustering loss; The expression for the total loss is as follows: In the formula, Indicates the reconstruction loss. Represents the clustering loss. For balance parameters; The expression for the reconstruction loss is: In the formula, Represents reconstructed features; Indicates key features, Represents all samples, Indicates the first Batch samples; The expression for the clustering loss is: In the formula, For the first The sample belongs to the first The soft assignment probability of clusters, The target distribution probability used for cluster optimization; The original probability distribution, which consists of all soft assignment probabilities, is used to describe the likelihood of a sample being assigned to a different cluster. The target probability distribution is formed by the probability distributions of all targets. It is deduced that this is used to improve the discriminative power and purity of clustering; express Divergence, used to quantify the distribution of a target Compared with the original distribution The differences between them; among them, and The calculation formula is as follows: In the formula, For the sample Embedded representation, For the first The cluster center vector corresponding to the cluster.

[0011] Preferably, the method includes using a multi-level particle swarm optimization algorithm to perform global hyperparameter optimization on the periodicity prediction model to obtain a trained periodicity prediction model, wherein the global hyperparameter optimization includes: Perform the first layer of particle swarm optimization. For the network structure parameters represented by each particle, call the optimal combination of hyperparameters obtained after the second layer of particle swarm optimization. Use the reciprocal of the decision variable as the fitness value to update the particle velocity and position to determine the optimal network structure parameters. The second layer of particle swarm optimization is performed. For each combination of hyperparameters represented by a particle, an initial cycle time prediction model is constructed based on the optimal network structure parameters after the first layer of particle swarm optimization. Particle iteration and inner layer iteration are then performed based on the fitness value and the initial cycle time prediction model to determine the optimal combination of hyperparameters.

[0012] Preferably, the step-by-step analysis of the importance of the multi-source data relative to the work-step cycle time based on the random forest model includes: For any node of each decision tree in the random forest, a single feature from the multi-source data is used to divide the current node into two child nodes, left and right, and the attenuation of the mean square error before and after the node split is calculated as the local gain of the current node. By traversing all decision trees of the random forest model, the global importance is obtained by weighted summing and averaging all local gains generated by the target feature at all nodes of all decision trees. The global importance is uniformly aggregated and normalized to obtain the importance of the target feature relative to the step-level cycle time.

[0013] Secondly, the present invention also provides a step-level cycle time prediction system for semiconductor wafer manufacturing, comprising: The data acquisition module acquires multi-source data throughout the entire semiconductor wafer manufacturing process; The feature extraction module analyzes the importance of the multi-source data relative to the step-level cycle time based on the random forest model, and selects key feature sets according to the degree of importance. The time prediction module processes the key feature set through a periodic time prediction model to obtain step-level periodic time prediction values. The periodic time prediction model includes a regression ensemble sub-model and a deep embedding clustering sub-model. The deep embedding clustering sub-model processes the key features to obtain an embedding representation and generates a soft clustering probability based on the embedding representation. The regression ensemble sub-model processes the embedding representation to obtain multiple local periodic time prediction values. The regression ensemble sub-model converts the soft clustering probability into clustering weights through a built-in dynamic modulation mechanism. The clustering weights and multiple local periodic time prediction values ​​are weighted and fused to generate the step-level periodic time prediction value.

[0014] Thirdly, this application also provides a computer-readable storage medium storing a computer program for step-level cycle time prediction in semiconductor wafer manufacturing, wherein the computer program causes a computer to execute the step-level cycle time prediction method for semiconductor wafer manufacturing as described above.

[0015] Fourthly, this application also provides an electronic device, comprising: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing step-level cycle time prediction methods for semiconductor wafer fabrication as described above.

[0016] (III) Beneficial Effects This invention provides a step-level cycle time prediction method and system for semiconductor wafer manufacturing. Compared with existing technologies, it has the following advantages: 1. This invention employs a cycle time prediction model composed of a deep embedding clustering sub-model and a regression ensemble sub-model. The deep embedding clustering sub-model extracts embedded representations from the selected key features and outputs the original cluster soft assignment probabilities. The regression ensemble sub-model generates multi-group local cycle time prediction values ​​based on the same embedded representation, and corrects the original cluster soft assignment probabilities through its built-in dynamic adjustment mechanism to generate dynamically adapted cluster weights. Finally, the cluster weights and multiple local cycle time prediction values ​​are weighted and fused to output the step-level cycle time prediction value. This significantly reduces the difference between the predicted cycle time and the actual wafer step cycle time, improves prediction accuracy from the underlying model structure level, overcomes the core technical pain points of traditional static weight prediction models that cannot adapt to the dynamic working conditions of the production line and have large prediction deviations, and greatly improves the practicality of this solution in actual semiconductor factory production scenarios.

[0017] 2. The deep embedding clustering sub-model of this invention mines the deep implicit correlations of multi-source process features of wafers through an autoencoder, compresses the original high-dimensional features to a low-dimensional embedding representation to filter irrelevant noise, and uses unsupervised clustering to distinguish wafer samples from different production conditions, accurately classifying sample clusters with different cycle time distributions; a local regression branch is configured separately for each type of cluster, unlike the global single regression model, which can match the unique cycle change patterns of samples from different production conditions and avoid systematic biases caused by global fitting; the dynamic adjustment mechanism built into the regression ensemble sub-model can iteratively correct the cluster membership weights in reverse according to the prediction error, breaking the limitation of fixed cluster allocation weights in existing schemes, and realizing bidirectional collaborative iterative optimization of cluster partitioning effect and cycle regression prediction results.

[0018] 3. This invention collects multi-source data from the entire semiconductor wafer manufacturing process, constructs a random forest model to quantify the importance of all features relative to the cycle time at the process step level, and selects key features strongly correlated with cycle time as input to the prediction model based on feature contribution. On the one hand, this significantly reduces the dimensionality of the model input, reduces the computational overhead caused by redundant data, accelerates the model training convergence speed, reduces the risk of overfitting caused by irrelevant features, and improves the model's generalization performance. On the other hand, the set of highly correlated process features obtained by screening has universality and can be used as a standardized feature reference for similar semiconductor wafer cycle prediction tasks, reducing the workload of data preprocessing development for similar prediction models. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A schematic diagram illustrating the steps of the step-level cycle time prediction method provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the structure of the periodic time prediction model provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] This application provides a method and system for predicting cycle time at the step level in semiconductor wafer manufacturing. It solves the problems of existing methods lacking real-time performance in CT (Cycle Time) prediction and failing to meet the requirements of precise step-level control, thereby achieving accurate prediction of step-level CT in the semiconductor wafer manufacturing process.

[0023] It's important to clarify that CT (Cut-off Time) in semiconductor manufacturing typically refers to the time required to process a batch of wafers (usually consisting of 25 wafers). A process step refers to the specific operation performed on the wafer within process modules such as photolithography and etching in the process route. These operations may be repeated during production. For example, a 300mm wafer typically requires 400 to 1000 process steps, and may be cycled 20 to 40 times in critical process stages to complete the layer-by-layer construction of the integrated circuit structure. (The term "batch" is a separate, unrelated concept and can be omitted.) Complete the steps The required time is defined as ,batch In the work steps The end time is defined as Then the batch during the production process .

[0024] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows: Most existing studies predict the total CT (Cut-off Dimension) of a wafer from its entry into the production line to its final shipment. However, wafer manufacturing, as a typical reentrant complex manufacturing system, usually involves hundreds or thousands of steps, each of which may undergo dynamic changes. Such CT predictions have significant limitations. Therefore, it is necessary to reasonably predict step-level CT to effectively ensure on-time product delivery.

[0025] Considering the challenges of modeling high-dimensional, multi-source, and dynamically changing data in step-level CT prediction, existing research typically employs a combined classification and prediction approach. This involves classifying batches with similar processing characteristics to improve the accuracy of the prediction model. For example, Deep Embedded Clustering (DEC) maps the original data to a low-dimensional embedding space and performs clustering optimization within this space. This achieves synergistic optimization of feature representation learning and data classification, providing a more reasonable structured input for subsequent CT prediction. Specifically, this method classifies batches based on representative features extracted from high-dimensional CT data that facilitate clustering, and then determines the clustering membership probability based on the samples. The DEC predicts step-level CT by weighted summation of the outputs of various regression models. However, DEC lacks a reasonable dynamic adjustment mechanism to update the results obtained from the clustering layer. This reduces the accuracy of predictions to some extent.

[0026] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0027] like Figure 1 As shown, this invention provides a step-level cycle time prediction method for semiconductor wafer manufacturing, comprising: Step S110: Obtain multi-source data for the entire semiconductor wafer manufacturing process; Step S120: Analyze the importance of the multi-source data relative to the work-step cycle time based on the random forest model, and select the key feature set according to the degree of importance; Step S130: The key feature set is processed by a cycle time prediction model to obtain step-level cycle time prediction values. The cycle time prediction model includes a regression ensemble sub-model and a deep embedding clustering sub-model. The deep embedding clustering sub-model processes the key features to obtain an embedding representation and generates a soft clustering probability based on the embedding representation. The regression ensemble sub-model processes the embedding representation to obtain multiple local cycle time prediction values. The regression ensemble sub-model converts the soft clustering probability into clustering weights through a built-in dynamic modulation mechanism. The clustering weights and multiple local cycle time prediction values ​​are weighted and fused to generate the step-level cycle time prediction values.

[0028] This invention employs a cycle time prediction model composed of a deep embedding clustering sub-model and a regression ensemble sub-model. The deep embedding clustering sub-model extracts embedded representations from the selected key features and outputs the original cluster soft assignment probabilities. The regression ensemble sub-model generates multi-group local cycle time prediction values ​​based on the same embedded representation and corrects the original cluster soft assignment probabilities through its built-in dynamic adjustment mechanism, generating dynamically adapted cluster weights. Finally, the cluster weights and multiple local cycle time prediction values ​​are weighted and fused to output the step-level cycle time prediction value. This significantly reduces the difference between the predicted cycle time and the actual wafer step cycle time, improving prediction accuracy from the underlying model structure level. It overcomes the core technical pain points of traditional static weight prediction models, which cannot adapt to the dynamic working conditions of the production line and have large prediction deviations, greatly improving the practicality of this solution in actual semiconductor factory production scenarios.

[0029] In one embodiment, step S110 involves acquiring multi-source data for the entire semiconductor wafer manufacturing process.

[0030] In this embodiment, multi-source data is collected throughout the entire semiconductor wafer manufacturing process. Preferably, data sources such as a Manufacturing Execution System (MES) are used. It is understood that multi-source data refers to various types of data collected from different systems, devices, or sensors during the semiconductor wafer manufacturing process. Examples include production batch information, equipment parameters, and process recipe data from the Manufacturing Execution System (MES), as well as environmental data and equipment status data collected by sensors.

[0031] In one embodiment, step S120 involves analyzing the importance of the multi-source data relative to the step-level cycle time based on a random forest model, and selecting a set of key features according to the degree of importance.

[0032] After obtaining multi-source data, features are constructed based on the multi-source data to obtain an initial feature set. The initial feature set includes batch priority features, design rule features, product layer features, batch type features, wafer quantity per batch features, remaining time until delivery features, work-in-process turnover features, completion rate features, main process group features, daily wafer output features, equipment group utilization features, equipment historical processing time features, equipment group historical waiting time features, and workshop shift features.

[0033] To improve the accuracy of the prediction results, this embodiment combines the analysis of the enterprise's actual production process and business needs, and selects relevant variables from data sources such as the MES (Manufacturing Execution System) to construct a feature set, as shown in Table 1. Table 1 lists the feature construction categories.

[0034] Table 1 Feature Construction Table Specifically, in the product and batch feature categories, the batch priority feature is used to determine the scheduling position of the batch in the queue processing process, and is divided into 1 to 6 levels, with lower priority being processed first; the design rule feature indicates the line width used in the semiconductor product design, and is divided into 3nm to 10nm; the layer number feature of each product indicates the number of layers required for product manufacturing, which will affect its priority in the processing process; the batch type feature specifically includes two types: mass production and R&D, and the batch type will affect the priority of the batch in the queue; the number of wafers per batch feature indicates the total number of wafers per batch, which is usually 25 wafers.

[0035] In the order feature category, the remaining time until delivery indicates the batch based on the original production plan. In the current work step The time remaining until its delivery deadline, i.e. ,in, and Each represents a batch In the work steps The remaining time and completion time, Indicates batch Delivery period; WIP (Work in Progress) turnaround: indicates the time required to meet batch delivery deadlines. The delivery deadline, the speed required to complete its remaining steps, i.e. ,in, Indicates batch In the work steps The speed required to complete the remaining steps Indicates batch The total number of steps in the product's process route; completion rate indicates the batch size. In the work steps At that time, the proportion of completed processing steps to the total number of steps, i.e. .

[0036] In the equipment status characteristic category, the main process group characteristics indicate that the wafer process steps can be divided into eight main process groups, including photolithography, etching, and metallization; daily wafer output ( Features represent batches Upon reaching the work step At that time, the average daily wafer output of the corresponding equipment group over the past 1 / 3 / 5 days; equipment group utilization rate ( Features represent batches Upon reaching the work step At that time, the equipment group utilization rate of the corresponding equipment group over the past hour / day / week is defined as follows: ,in , They represent time respectively ( Internal equipment Actual and planned operating times; historical processing time of equipment group ( ) and waiting time ( Features represent batches Upon reaching the work step At that time, the historical processing time and waiting time of all jobs processed by the corresponding equipment group within the past hour / day / week are recorded, and the statistical indicators include the minimum value ( ), maximum value ( ), mean ( ) and variance ( ).

[0037] In the workshop condition characteristic category, the shift characteristic indicates that semiconductor workshop shifts can be divided into weekends, holidays, and weekdays according to date, and into day shifts (6:00-14:00), afternoon shifts (14:00-22:00), and night shifts (22:00-6:00) according to time; batches Upon reaching the work step The specific schedule will be determined based on the time.

[0038] After obtaining the feature set, the importance of the feature set to the step-level cycle time in the multi-source data is analyzed by using a random forest model, and key features are selected according to their importance.

[0039] Understandably, key features refer to those features that have a significant impact on cycle time at the step level, selected from raw multi-source data through feature importance analysis. These features are typically variables closely related to production efficiency, equipment status, and process parameters.

[0040] Random forests, based on an ensemble learning framework, evaluate features by constructing a forest model composed of numerous regression decision trees. This effectively quantifies the explanatory power of features for the prediction target, thereby selecting highly relevant features. Its core algorithmic logic lies in calculating the mean squared error decay of each feature during the splitting process of a decision tree node to measure the contribution of each feature to the improvement of node purity, thus quantifying its importance to the prediction target. For any node in a single regression decision tree... The formula for calculating the mean square error of the target value of its internal samples is: in, To reach the node The total number of samples, To reach the node The sample set, For the sample Actual CT values ​​at the step level, To reach the node The mean of all target values ​​for the samples. Based on this, the specific calculation process of random forest regression includes steps S121~S123.

[0041] Step S121: For any node of each decision tree in the random forest, use a single feature from the multi-source data to divide the current node into left and right child nodes, and calculate the attenuation of the mean square error before and after the node split as the local gain of the current node.

[0042] Specifically, the local gain is first calculated, that is, the calculation model is based on the features. Node Divided into left child nodes and right child node The resulting error attenuation; its calculation formula is as follows: in, It uses features Parent node Error decay after performing the splitting operation This represents the mean squared error. Understandably, Indicates the parent node to be split; , These represent the left and right child nodes after the split, respectively. , , Representing nodes respectively , , Total number of samples included.

[0043] Step S122: Traverse all decision trees of the random forest model, and sum the weighted local gains generated by the target feature at all nodes of all decision trees and then take the average value to obtain the global importance.

[0044] Then the features are calculated. global importance To eliminate the randomness of a single decision tree, it is defined as a feature. All in the forest The weighted sum of the error attenuation contributed by each decision tree during the split; its calculation formula is: in, Indicates the first Among the trees, the characteristics The set of nodes to be split. Represents a set Any execution feature Split node index, For the first The total number of training samples corresponding to each decision tree. For training the first Total number of samples per node The total number of decision trees. This represents the global importance of a single binary derived feature after one-hot encoding splitting.

[0045] Step S123: Perform uniform aggregation and normalization on the global importance values ​​to obtain the importance of the target feature relative to the step-level cycle time.

[0046] Aggregate feature importance: Since categorical variable features are split into multiple binary variables using one-hot encoding, the actual contribution of the original features is considered. By analyzing all derived features belonging to the same original feature The importance of elements is aggregated and reconstructed; the calculation formula is: In the formula, To reconstruct the complete global importance of the original classification features after aggregating the importance of all homologous binary derived features; The total number of binary derived features obtained by splitting a single original classification feature using one-hot encoding. For derived feature indexes.

[0047] Finally, the features are processed by normalization. After obtaining the value, the relative importance percentage of each feature relative to the prediction target can be obtained.

[0048] To accurately and reliably assess feature importance, this embodiment utilizes the scikit-learn library in a Python environment to construct a random forest (RF) regression model. The experiment employs a multi-core parallel computing mode, utilizing all available CPU cores to construct the decision trees in the forest in parallel, thereby improving computational efficiency. Based on preliminary experimental results, the number of decision trees in the random forest was set to 100, and the random seed was set to 42. After calculating the importance percentage of all original features, they were sorted from highest to lowest importance, and the top 17 important features with a cumulative contribution rate exceeding 85% were selected for prediction. The final selected key features are shown in the table below: Table 2 Key Feature Set As shown in Table 2, the key feature set selected in this embodiment includes the main process group features. Features of each product level Daily wafer production characteristics (the average daily wafer output of the corresponding equipment group over the past day at this moment) Average daily wafer output over the past 3 days Average daily wafer output over the past 5 days ), Historical processing time characteristics of equipment group (the minimum historical processing time of all jobs processed by the corresponding equipment group in the past hour at this moment). Maximum historical processing time Historical average processing time Historical processing time variance The minimum historical processing time of all jobs processed by the corresponding equipment group within the past day at this moment. Maximum historical processing time Historical average processing time Historical processing time variance The minimum historical processing time for all jobs processed by the corresponding equipment group within the past week at this moment. Maximum historical processing time Historical average processing time Historical processing time variance Batch type characteristics Characteristics of the number of wafers per batch .

[0049] In one embodiment, step S130 involves processing the key features using a pre-trained and fine-tuned cycle time prediction model to obtain step-level cycle time prediction values. The cycle time prediction model includes a regression ensemble sub-model and a deep embedding clustering sub-model. The deep embedding clustering sub-model processes the key features to obtain an embedding representation and generates a soft clustering probability based on the embedding representation. The regression ensemble sub-model processes the embedding representation to obtain multiple local cycle time prediction values. The regression ensemble sub-model converts the soft clustering probability into clustering weights through a built-in dynamic modulation mechanism. The clustering weights and multiple local cycle time prediction values ​​are weighted and fused to generate the step-level cycle time prediction values.

[0050] The deep embedding clustering sub-model processes the key features to obtain embedding representations and generates clustering soft assignment probabilities based on the embedding representations, including: Step S131: Encode the key features using an autoencoder to obtain an embedded representation. ; Step S132: Process the embedded representation through a clustering layer to obtain the clustering soft assignment probability.

[0051] Specifically, after obtaining the key features, they are input into the periodic time prediction model. The periodic time prediction model first uses an autoencoder in a deep embedding clustering sub-model to process the key features ( Encoding as an embedded representation The calculation formula is as follows: In the formula, Represents the weight matrix. Indicates encoder bias. This represents the activation function.

[0052] The embedding representation is then processed through a clustering layer in the deep embedding clustering sub-model. Obtain the cluster soft assignment probability .

[0053] See Figure 2 , Figure 2The Deep Embedded Clustering (DEC) pre-training stage in the diagram demonstrates the structure of the deep embedded clustering sub-model, which consists of a stacked autoencoder (AE) and clustering layers. It can simultaneously learn feature representations and cluster assignments using a deep neural network. Specifically, the AE includes a decoder and an encoder, which can... A dataset consisting of samples Mapping to low-dimensional feature representation and will Reconstructing the dataset Furthermore, the encoded feature representation The data is input into the clustering layer for soft clustering, and the soft clustering probability is obtained.

[0054] See you again Figure 2 , Figure 2 The DEC (fine-tuning) part of the combined regression ensemble network demonstrates the application of the regression ensemble sub-model. The regression ensemble sub-model establishes a regression network for multiple clusters in the clustering layer and represents it in low dimension. As input, for each sample The regression network outputs local periodic time predictions. Subsequently, the output of the clustering layer is adjusted through the dynamic modulation mechanism built into the regression ensemble sub-model. Obtain cluster weights The specific implementation steps include: Step S410, Calculate the adjustment amount: Use the Multi-Layer Perceptron (MLP) in the regression ensemble sub-model based on the samples. The original cluster soft assignment probability Calculate the correction vector The formula for calculating the correction vector is: In the formula, , This represents the weights and biases of the first layer of the network; , This represents the weights and biases of the second layer network.

[0055] Step S420: Add the cluster soft assignment probability to the correction vector.

[0056] Step S430: Renormalize and correct the vector using the Softmax function. Cluster soft assignment probability The sum of the values ​​yields the cluster weights. The calculation formula is as follows: Furthermore, the first The output values ​​of each cluster The calculation formula is: in, For the correction vector The first in Each element.

[0057] Finally, the clustering weights and multiple local periodic time predictions are weighted and fused to generate prediction samples. The predicted cycle time for each work step is calculated using the following formula: In the formula, Indicates the first Batch samples, The cluster index represents the first cluster index. A cluster, This represents the total number of clusters in the clustering layer. Indicates the first Predicted batch sample cycle time.

[0058] Next, we will introduce the training and fine-tuning process of the periodic time prediction model.

[0059] The deep embedding clustering sub-model includes a pre-training phase, which specifically includes the following implementation steps: Step S210: Obtain the key feature set; preset the balance parameters and number of clusters for the deep embedding clustering sub-model. ; Initialize the hyperparameters of the deep embedding clustering sub-model; Initialize the encoder weights and decoder weights of the stacked autoencoders in the deep embedding clustering sub-model; Step S220: Process the key feature set through the stacked autoencoder to obtain an embedded representation; Step S230: Cluster the embedding representation using a clustering algorithm to initialize... Cluster centers ; Step S240: Calculate the total loss based on the balance parameters, the current embedding representation, and the cluster centers; Step S250: Based on the total loss, update the cluster centers, encoder weights and decoder weights in reverse using the gradient descent algorithm, and obtain a new embedding representation based on the updated stacked autoencoder; Step S260: Repeat steps S240 to S250 until the maximum number of iterations is reached to obtain the parameters of the pre-trained deep embedding clustering sub-model.

[0060] It should be clarified that the input content of the deep embedding clustering pre-training process in this embodiment includes: key feature set. Loss balance parameters Preset number of clusters The maximum number of training iterations is specified; the output consists of the pre-trained cluster centers, encoder weights, and decoder weights.

[0061] Specifically, 1) Initialize all hyperparameters of the Deep Embedded Clustering Network (DEC). These hyperparameters are the optimal combination output from the two-layer MPSO hyperparameter optimization process, including topological parameters representing the depth and width of the encoder, as well as numerical hyperparameters such as learning rate, batch size, dropout rate, and number of training epochs, thus completing the complete network structure of the DEC autoencoder. 2) Perform unsupervised pre-training on the DEC stacked autoencoder separately to complete the basic weights of the encoder. Decoder base weights The initialization enables the autoencoder to initially map high-dimensional process features to a low-dimensional embedding space. 3) The key features are encoded by the encoder in the autoencoder to obtain the embedding representation, and the embedding representation is processed by the K-means clustering algorithm to initialize the embedding space. Cluster centers 4) Initialize the number of iterations. 5) Based on the weight parameters of the current encoder and decoder in this round, calculate the reconstruction loss corresponding to the autoencoder. 6) Feed the current batch of samples into the encoder network and solve for the low-dimensional embedding vector corresponding to each sample. Soft assignment probability and target distribution 7) Calculate the total loss according to the preset formula. 8) With minimizing the total loss as the optimization objective, perform backpropagation update using stochastic gradient descent, and update the cluster centers synchronously. Encoder weights and decoder weights Three types of parameters. 9) Execute iteration steps 5-7 until the preset maximum number of iterations is reached, then terminate the iteration. 10) Output the cluster centers of the trained and optimized parameters. Encoder weights and decoder weights .

[0062] In this embodiment, the total loss for training and fine-tuning the periodic time prediction model includes reconstruction loss and clustering loss; wherein, the expression for the total loss is: In the formula, Indicates the reconstruction loss. Represents the clustering loss. These are the balancing parameters.

[0063] The reconstruction loss is obtained by encoding the key features to obtain an embedded representation, and then decoding the embedded representation to obtain the reconstructed features. Then calculate the reconstructed features and key features The error is obtained; where the embedding representation is... Decode into data The expression is: In the formula, Represents the weight matrix. This indicates the bias of the decoder.

[0064] The expression for the reconstruction loss is: In the formula, Represents reconstructed features; Indicates key features, Represents all samples, Indicates the first Batch samples.

[0065] The expression for the clustering loss is: in, For the first The sample belongs to the first The soft assignment probability of clusters, The target distribution probability used for cluster optimization; The original probability distribution, which consists of all soft assignment probabilities, is used to describe the likelihood of a sample being assigned to a different cluster. The target probability distribution is formed by the probability distributions of all targets. It is deduced that this is used to improve the discriminative power and purity of clustering; express Divergence, used to quantify the distribution of a target Compared with the original distribution The differences between them. Among them, and The calculation formula is as follows: In the formula, For the sample Embedded representation, For the first The cluster center vector corresponding to the cluster. According to each cluster pair The frequency after squaring is normalized.

[0066] After pre-training the deep embedding clustering sub-model, the periodic time prediction model is fine-tuned. The specific implementation method includes the following steps: Step S310: Obtain the embedding representation and soft assignment probability output by the deep embedding clustering sub-model; Step S320: Remove the decoder part of the deep embedding clustering sub-model and construct... A regression subnetwork is created and connected to the encoder; Step S330: Initialize the weights of the regression sub-networks and the parameters of the multilayer perceptron in the regression ensemble sub-model; Step S340: The embedded representation and soft-assignment probability are processed by a regression ensemble sub-model to obtain the predicted cycle time value, and the prediction loss is calculated based on the predicted cycle time value and the actual cycle time value. Step S350: Based on the prediction loss, the cluster centers, autoencoder weights, regressor subnetwork weights, and multilayer perceptron parameters are updated in reverse using the gradient descent algorithm to obtain the fine-tuned periodic time prediction model.

[0067] In this embodiment, after pre-training the embedded model to obtain parameters, a clustered regression ensemble network is built on the basis of the pre-trained encoder and cluster centers. Joint fine-tuning of clustering, encoding, and cycle time prediction is carried out, and finally, a complete set of optimized model parameters that can be used for wafer process cycle time inference is output.

[0068] It needs to be clarified that the input to the fine-tuning process includes: the set of low-dimensional embedding feature representations of all samples obtained by mapping through the pre-trained encoder. The set consists of the low-dimensional embedding vector corresponding to each sample. Composition; actual cycle time obtained from each step of the wafer fabrication process. Cluster centers obtained from pre-training of a deep clustering model Encoder weights The output includes: cluster centers after joint fine-tuning. DEC encoder weights , The weight parameters corresponding to each clustered regression network MLP parameters; where the MLP parameters consist of two layers of weight matrices. , and the corresponding bias terms of the two layers , .

[0069] Specifically, 1) Remove the decoder structure from the pre-trained DEC model and build a new structure equal to the number of clusters. Each regression subnetwork is an independent regression subnetwork, and the input ports of all regression subnetworks are connected to the pre-trained encoder. The decoder in the pre-training stage is only used for reconstruction of the original process features. It does not perform reconstruction tasks in the fine-tuning stage, so the decoder is directly removed. Each regression subnetwork corresponds to a wafer production condition cluster, specifically fitting the variation law of the cycle time of the process step under a single condition. The low-dimensional embedding vector output by the encoder is used as the shared input data of all regression subnetworks. 2) Initialize the weights of the regression subnetworks. and MLP parameters, where the MLP parameters consist of two layers of weight matrices. , and the corresponding bias terms of the two layers , 3) Initialize the iteration number. 4) Represent the set using the low-dimensional features corresponding to the samples in this batch. For input, the first The low-dimensional embedding vectors of each sample are respectively fed into The nth independent regression subnetwork is calculated to obtain the nth... The sample in the first The corresponding periodic time prediction output value on the regression network 5) Obtain the independent prediction results of the regression subnetwork matching all working conditions for a single sample. 6) Activate the dynamic adjustment mechanism based on a two-layer MLP to convert the soft-assignment probability of clustering into cluster weights. 7) Weight and fuse the cluster weights and the corresponding cycle time predictions to generate the cycle time predictions at the work step level. 7) Update the cluster centers, encoder weights, regressor submodel weights, and multilayer perceptron parameters using stochastic gradient descent. 8) Iterate through steps 4-7 until the preset maximum number of iterations is reached. 9) Output the cluster centers. DEC encoder weights , The weight parameters corresponding to each clustered regression network MLP parameters.

[0070] In this embodiment, the loss is predicted. Calculated as predicted value and actual value The comparison is calculated using the following formula: The prediction error will be fed back to the dynamic adjustment mechanism and the clustering layer respectively, and the clustering soft assignment probability will be updated iteratively to achieve synergistic optimization of the clustering results and the prediction model, thereby improving the accuracy of the prediction results.

[0071] In this embodiment, a multi-level particle swarm optimization algorithm is used to perform global hyperparameter optimization on the periodic time prediction model. The global hyperparameter optimization includes: Step S510: Perform the first layer of particle swarm optimization. For the network structure parameters represented by each particle, call the optimal hyperparameter combination obtained after the second layer of particle swarm optimization. Use the reciprocal of the decision variable as the fitness value to update the particle velocity and position to determine the optimal network structure parameters.

[0072] Specifically, the input to the first-layer particle swarm optimization algorithm is: particle swarm size. Maximum number of iterations The output is: the optimal number of coding layers. and regression network layers The specific implementation steps are as follows: Step S511, Initialize outer particle swarm parameters: Set the outer particle swarm size and maximum number of iterations For each particle in the swarm Randomly initialize its position vector ,in The number of layers in the DEC encoder structure. The number of layers in the regression network structure. Initialize the number of iterations. .

[0073] Step S512: Determine whether the current iteration counter satisfies the condition. If satisfied, proceed to steps S513 to S518 sequentially; if not satisfied, proceed to step S519.

[0074] Step S513: Calculate the inertia weight for the current iteration according to the following formula. : ; Step S514: Iterate through each particle in the swarm. Then, proceed with steps S515 to S517.

[0075] Step S515, set the current particle position The network structure parameters represented ( The result is passed to the second-layer particle swarm optimization algorithm, which then performs inner-layer particle swarm optimization under the current network structure parameter constraints to search for the optimal hyperparameter combination for the corresponding network structure. The minimum fitness value returned by the second-layer particle swarm optimization algorithm is received as the current particle position. fitness value .

[0076] Step S516, based on fitness value Update particles Individual historical best position and the global optimal position of the group .

[0077] Step S517: Based on the particle swarm optimization velocity update formula and position update formula, and using the inertia weights calculated in step S513, update the particle... speed and location .

[0078] Step S518, update the iteration number Then return to step S512.

[0079] Step S519: Output the optimal number of coding layers. and regression network layers .

[0080] Step S520: Perform the second layer of particle swarm optimization. For each combination of hyperparameters represented by a particle, construct an initial cycle time prediction model based on the optimal network structure parameters after the first layer of particle swarm optimization. Then, perform particle iteration and inner layer iteration based on the fitness value and the initial cycle time prediction model to determine the optimal combination of hyperparameters.

[0081] Particle position vector Used to characterize network structure parameters, it can be understood that the parameters In the perspective of particle swarm optimization, it refers to the particle position. When particle swarm optimization is applied to optimize model parameters, the "solution" sought by particle swarm optimization is itself the network structure parameters. The two are mapped. The position vector is the "encoding / carrier", and the network structure parameters are the "content / semantics" it carries. Every time a particle moves to a new coordinate, it is equivalent to proposing a new set of network structure schemes.

[0082] Specifically, the implementation steps of the second-layer particle swarm optimization algorithm are as follows: Step S521: Receive the current network structure parameters passed from the first-layer particle swarm optimization algorithm. Set the inner particle swarm size and maximum inner iteration number Among them, the current network structure parameters Including the optimal number of coding layers and regression network layers .

[0083] Step S522, iterate through each particle in the swarm. Perform steps S523 to S524.

[0084] Step S523, based on Random initialization of the first The position vector of each particle Position vector Including the dimensions of each hidden layer of the DEC encoder Dimensions of each hidden layer in the regression network Learning rate ( ); Drop rate ( Batch size () Training rounds () ); number of clusters ( ); Balance parameters ( ).

[0085] Step S524, based on and Construct a DEC regression network; and calculate the fitness value based on the constructed network. Based on fitness values Update individual historical best position and the global optimal position of the group .

[0086] Step S525, initialize the number of iterations. .

[0087] Step S526, determine Is it true? If it is true, proceed to steps S527 to S533 in sequence; if it is false, proceed to step S534.

[0088] Step S527, update the number of iterations. .

[0089] Step S528: Calculate the inertia weight for the current iteration according to the following formula. : ; Step S529, traverse the particles Perform steps S530 to S532.

[0090] Step S530: Based on the particle swarm optimization velocity update formula and position update formula, use the inertia weights calculated in step S528. Update # particle speed and location .

[0091] Step S531, based on the updated position vector The hyperparameter combination represented, and the fixed outer structure parameters Reconstruct the DEC regression network model and calculate the new fitness value. .

[0092] Step S532, based on the new fitness value Update # The optimal position of each particle and the global optimal position of the group .

[0093] Step S533, return to step S526.

[0094] Step S534: Output the optimal combination of hyperparameters. And fitness value.

[0095] Furthermore, when the determination coefficient corresponding to the validation dataset is lower than the first preset threshold, the fitness value is corrected to a preset penalty constant; when the difference between the determination coefficient corresponding to the training dataset and the determination coefficient corresponding to the validation dataset exceeds the second preset threshold, the fitness value is corrected by adding a preset penalty increment to the basic fitness value.

[0096] This embodiment adjusts the hyperparameters of the prediction model based on the Multilevel Particle Swarm Optimization (MPSO) algorithm. Its core logic is as follows: First, the MPSO algorithm randomly generates a set of hyperparameters in each iteration; then, based on this set of hyperparameters, the model is trained using 70% of the dataset and uses the remaining 30% of the dataset for prediction to obtain the coefficient of determination. reciprocal .at last, The fitness value is fed back to the MPSO algorithm to obtain a better combination of hyperparameters. The purpose of this is to maximize... .

[0097] In addition, in order to prevent Furthermore, we aim to find hyperparameters that can keep it at least 0.6. This embodiment sets up the following penalty mechanism: when At that time, the fitness was 1000; when Add 100 to the original fitness level.

[0098] The hyperparameters of the model in this study include: DEC encoder structure; regression network structure; learning rate (…). ); Drop rate ( Batch size () Training rounds () ); number of clusters ( ); Balance parameters ( The adjustment ranges for each hyperparameter are as follows: The DEC encoder structure has layer, The integers are in the range [2, 4]. Since the dimension of the input data is... The DEC encoder structure can be represented as follows: .in, For low-dimensional embedding representation The dimension, with a value range of . Integers; Indicates the first The dimension of the layer The range of values ​​is integers, ( The range of values ​​is Integers.

[0099] Regression network structure has layer, For integers in the range [1, 3], its structure can be represented as follows: .in, Indicates the first The dimension of the layer The range of values ​​is integers, ( The range of values ​​is Integers.

[0100] , and They are respectively and A random decimal. and They are respectively and A random integer. Set as , for A random integer.

[0101] The MPSO algorithm first performs a first-level particle swarm optimization to optimize the network structure parameters. and Then, under this layer configuration, a second layer of particle swarm optimization is performed to optimize the specific parameters of each layer.

[0102] This invention also provides a step-level cycle time prediction system for semiconductor wafer manufacturing, comprising: The data acquisition module acquires multi-source data throughout the entire semiconductor wafer manufacturing process; The feature extraction module analyzes the importance of the multi-source data relative to the step-level cycle time based on the random forest model, and selects key feature sets according to the degree of importance. The time prediction module processes the key feature set through a periodic time prediction model to obtain step-level periodic time prediction values. The periodic time prediction model includes a regression ensemble sub-model and a deep embedding clustering sub-model. The deep embedding clustering sub-model processes the key features to obtain an embedding representation and generates a soft clustering probability based on the embedding representation. The regression ensemble sub-model processes the embedding representation to obtain multiple local periodic time prediction values. The regression ensemble sub-model converts the soft clustering probability into clustering weights through a built-in dynamic modulation mechanism. The clustering weights and multiple local periodic time prediction values ​​are weighted and fused to generate the step-level periodic time prediction value.

[0103] It is understood that the step-level cycle time prediction system for semiconductor wafer manufacturing provided in this embodiment of the invention corresponds to the step-level cycle time prediction method for semiconductor wafer manufacturing described above. The explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding content in the step-level cycle time prediction method for semiconductor wafer manufacturing, and will not be repeated here.

[0104] This invention also provides a computer-readable storage medium storing a computer program for a step-level cycle time prediction method for semiconductor wafer manufacturing, wherein the computer program causes a computer to execute the step-level cycle time prediction method for semiconductor wafer manufacturing as described above.

[0105] This application also provides an electronic device, including: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing step-level cycle time prediction methods for semiconductor wafer manufacturing as described above.

[0106] In summary, compared with existing technologies, it has the following beneficial effects: 1. This invention employs a cycle time prediction model composed of a deep embedding clustering sub-model and a regression ensemble sub-model. The deep embedding clustering sub-model extracts embedded representations from the selected key features and outputs the original cluster soft assignment probabilities. The regression ensemble sub-model generates multi-group local cycle time prediction values ​​based on the same embedded representation, and corrects the original cluster soft assignment probabilities through its built-in dynamic adjustment mechanism to generate dynamically adapted cluster weights. Finally, the cluster weights and multiple local cycle time prediction values ​​are weighted and fused to output the step-level cycle time prediction value. This significantly reduces the difference between the predicted cycle time and the actual wafer step cycle time, improves prediction accuracy from the underlying model structure level, overcomes the core technical pain points of traditional static weight prediction models that cannot adapt to the dynamic working conditions of the production line and have large prediction deviations, and greatly improves the practicality of this solution in actual semiconductor factory production scenarios.

[0107] 2. The deep embedding clustering sub-model of this invention mines the deep implicit correlations of multi-source process features of wafers through an autoencoder, compresses the original high-dimensional features to a low-dimensional embedding representation to filter irrelevant noise, and uses unsupervised clustering to distinguish wafer samples from different production conditions, accurately classifying sample clusters with different cycle time distributions; a local regression branch is configured separately for each type of cluster, unlike the global single regression model, which can match the unique cycle change patterns of samples from different production conditions and avoid systematic biases caused by global fitting; the dynamic adjustment mechanism built into the regression ensemble sub-model can iteratively correct the cluster membership weights in reverse according to the prediction error, breaking the limitation of fixed cluster allocation weights in existing schemes, and realizing bidirectional collaborative iterative optimization of cluster partitioning effect and cycle regression prediction results.

[0108] 3. In this embodiment of the invention, multi-source data from the entire semiconductor wafer manufacturing process are collected, and a random forest model is constructed to quantify the importance of all features relative to the cycle time at the process step level. Based on the feature contribution, key features strongly correlated with cycle time are selected as inputs to the prediction model. On the one hand, this significantly reduces the dimensionality of the model input, reduces the computational overhead caused by redundant data, accelerates the model training convergence speed, reduces the risk of overfitting caused by irrelevant features, and improves the model's generalization performance. On the other hand, the set of highly correlated process features obtained by selection has universality and can be used as a standardized feature reference for similar semiconductor wafer cycle prediction tasks, reducing the workload of data preprocessing development for similar prediction models.

[0109] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0110] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting cycle time at the step level in semiconductor wafer manufacturing, characterized in that, include: Acquire multi-source data throughout the entire semiconductor wafer manufacturing process; The importance of the multi-source data to the step-level cycle time was analyzed based on the random forest model, and key feature sets were selected according to the degree of importance. The key feature set is processed by a cycle time prediction model to obtain step-level cycle time prediction values. The cycle time prediction model includes a regression ensemble sub-model and a deep embedding clustering sub-model. The deep embedding clustering sub-model processes the key features to obtain an embedding representation and generates a soft clustering probability based on the embedding representation. The regression ensemble sub-model processes the embedding representation to obtain multiple local cycle time prediction values. The regression ensemble sub-model converts the soft clustering probability into clustering weights through a built-in dynamic modulation mechanism. The clustering weights and multiple local cycle time prediction values ​​are weighted and fused to generate the step-level cycle time prediction values.

2. The method for predicting the cycle time at the step level according to claim 1, characterized in that, The regression ensemble sub-model converts the clustering soft assignment probabilities into clustering weights through a built-in dynamic modulation mechanism, including: The correction vector is calculated using the multilayer perceptron in the regression ensemble sub-model based on the cluster soft assignment probability. Add the clustering soft assignment probability and the correction vector; The clustering weights are obtained by renormalizing the clustering soft assignment probabilities and the sum of the correction vectors using a vector-level normalized activation function.

3. The method for predicting the cycle time at the step level according to claim 1, characterized in that, The deep embedding clustering sub-model includes a pre-training phase, which includes: Step S210: Obtain the key feature set; preset the balance parameters and number of clusters for the deep embedding clustering sub-model. Initialize the encoder weights and decoder weights of the stacked autoencoders in the deep embedding clustering sub-model; Step S220: Process the key feature set through the stacked autoencoder to obtain an embedded representation; Step S230: Cluster the embedding representation using a clustering algorithm to initialize... Cluster centers ; Step S240: Calculate the total loss based on the balance parameters, the current embedding representation, and the cluster centers; Step S250: Based on the total loss, update the cluster centers, encoder weights and decoder weights in reverse using the gradient descent algorithm, and obtain a new embedding representation based on the updated stacked autoencoder; Step S260: Repeat steps S240 to S250 until the maximum number of iterations is reached to obtain the parameters of the pre-trained deep embedding clustering sub-model.

4. The method for predicting the cycle time at the step level according to claim 3, characterized in that, After pre-training the deep embedding clustering sub-model, the following is included: Obtain the embedding representation and cluster soft assignment probability output by the deep embedding clustering sub-model; Remove the decoder part of the deep embedding clustering sub-model and construct A regression subnetwork is created and connected to the encoder; Initialize the weights of the regression sub-networks and the parameters of the multilayer perceptron in the regression ensemble sub-model; The periodic time prediction is obtained by processing the embedded representation and soft-assigned probability through a regression ensemble sub-model, and the prediction loss is calculated based on the periodic time prediction and the actual periodic time. Based on the predicted loss, the cluster centers, autoencoder weights, regressor subnetwork weights, and multilayer perceptron parameters are updated in reverse using the gradient descent algorithm to obtain a fine-tuned periodic time prediction model.

5. The method for predicting the cycle time at the step level according to claim 3, characterized in that, The total loss includes reconstruction loss and clustering loss; The expression for the total loss is as follows: In the formula, Indicates the reconstruction loss. Represents the clustering loss. For balance parameters; The expression for the reconstruction loss is: In the formula, Represents reconstructed features; Indicate key features; Represents all samples, Indicates the first Batch samples; The expression for the clustering loss is: In the formula, For the first The sample belongs to the first The soft assignment probability of clusters, The target distribution probability used for cluster optimization; The original probability distribution, which consists of all soft assignment probabilities, is used to describe the likelihood of a sample being assigned to a different cluster. The target probability distribution is formed by the probability distributions of all targets. It is deduced that this is used to improve the discriminative power and purity of clustering; express Divergence, used to quantify the distribution of a target Compared with the original distribution The differences between them; among them, and The calculation formula is as follows: In the formula, For the sample Embedded representation, For the first The cluster center vector corresponding to the cluster.

6. The method for predicting the cycle time at the step level according to claim 1, characterized in that, It also includes using a multi-level particle swarm optimization algorithm to perform global hyperparameter optimization on the period-time prediction model to obtain the trained period-time prediction model, wherein the global hyperparameter optimization includes: Perform the first layer of particle swarm optimization. For the network structure parameters represented by each particle, call the optimal combination of hyperparameters obtained after the second layer of particle swarm optimization. Use the reciprocal of the decision variable as the fitness value to update the particle velocity and position to determine the optimal network structure parameters. The second layer of particle swarm optimization is performed. For each combination of hyperparameters represented by a particle, an initial cycle time prediction model is constructed based on the optimal network structure parameters after the first layer of particle swarm optimization. Particle iteration and inner layer iteration are then performed based on the fitness value and the initial cycle time prediction model to determine the optimal combination of hyperparameters.

7. The method for predicting the cycle time at the step level according to claim 1, characterized in that, The analysis of the importance of the multi-source data relative to the step-level cycle time based on the random forest model includes: For any node of each decision tree in the random forest, a single feature from the multi-source data is used to divide the current node into two child nodes, left and right, and the attenuation of the mean square error before and after the node split is calculated as the local gain of the current node. By traversing all decision trees of the random forest model, the global importance is obtained by weighted summing and averaging all local gains generated by the target feature at all nodes of all decision trees. The global importance is uniformly aggregated and normalized to obtain the importance of the target feature relative to the step-level cycle time.

8. A step-level cycle time prediction system for semiconductor wafer manufacturing, characterized in that, include: The data acquisition module acquires multi-source data throughout the entire semiconductor wafer manufacturing process; The feature extraction module analyzes the importance of the multi-source data relative to the step-level cycle time based on the random forest model, and selects key feature sets according to the degree of importance. The time prediction module processes the key feature set through a periodic time prediction model to obtain step-level periodic time prediction values. The periodic time prediction model includes a regression ensemble sub-model and a deep embedding clustering sub-model. The deep embedding clustering sub-model processes the key features to obtain an embedding representation and generates a soft clustering probability based on the embedding representation. The regression ensemble sub-model processes the embedding representation to obtain multiple local periodic time prediction values. The regression ensemble sub-model converts the soft clustering probability into clustering weights through a built-in dynamic modulation mechanism. The clustering weights and multiple local periodic time prediction values ​​are weighted and fused to generate the step-level periodic time prediction value.

9. A computer-readable storage medium, characterized in that, It stores a computer program for predicting step-level cycle time for semiconductor wafer manufacturing, wherein the computer program causes a computer to execute the step-level cycle time prediction method for semiconductor wafer manufacturing as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing step-level cycle time prediction methods for semiconductor wafer fabrication as described in any one of claims 1 to 7.

Images