Semiconductor manufacturing intelligent prediction closed-loop control method and system and readable storage medium

Abstract

The invention discloses an intelligent prediction closed-loop control method and system for semiconductor manufacturing and a readable storage medium, and the method comprises the following steps: obtaining real-time FDC data of a machine, inputting the FDC data into a preset virtual measurement model, and outputting a first Inline measurement parameter of a product in process; inputting the third Inline measurement parameter into a preset Inline-Wat prediction model, and outputting to obtain a predicted Wat value of the product in process; and when it is judged that the predicted Wat value exceeds an abnormal control range, the Inline measurement parameters of the current station process and the later station process of the product are corrected. According to the semiconductor manufacturing intelligent prediction closed-loop control method and system and the readable storage medium, high-precision prediction, real-time monitoring and process parameter automatic correction in the semiconductor manufacturing process are achieved, the prediction precision and reliability are improved, and the efficient prediction capacity is kept.

Description

Technical Field

[0001] The present invention relates to the technical field of semiconductor manufacturing, and particularly to an intelligent prediction closed-loop control method, system and readable storage medium for semiconductor manufacturing. Background Art

[0002] In the field of advanced semiconductor manufacturing processes, as the chip feature size evolves towards 3 nanometers and below nodes, significant multi-physical field coupling effects emerge in the multi-process collaborative manufacturing process. Currently, the production line integrates more than 500 process steps, and the high-order non-linear interaction between process parameters grows exponentially, posing unprecedented challenges to the precise control of the manufacturing process. Among them, Inline Wat (Wafer Acceptable Test) is a crucial concept in semiconductor manufacturing. It refers to the Wat test carried out at a certain stage during wafer manufacturing (such as the inter-metal stage). The purpose of the test is to check the electrical performance parameters of the wafer in real time during the process, monitor the quality of the wafer during production, ensure that the electrical characteristics of each process step meet the design requirements, and thus ensure the yield of the finished product.

[0003] Traditional process modeling methods based on expert experience adopt linear modeling methods based on simplified rules, and expose systematic defects when dealing with the following core problems: insufficient ability to characterize complex non-linear features: there is a strong coupling non-linear mapping relationship between key process indicators (such as dielectric deposition thickness, lithography critical dimension CD) and Wat electrical parameters. Experimental data shows that the capture error of traditional empirical models for high-order interaction terms and time-varying features exceeds 35%, resulting in the root mean square error (RMSE) between the predicted value and the measured value of the process window generally being higher than the industry accuracy threshold, and it is difficult to support the process control requirements at the sub-nanometer level.

[0004] Lack of dynamic process adaptive mechanism: In the industrial background where the iteration cycle of advanced materials such as new high-k dielectrics and EUV photoresists is shortened to 6-8 months, traditional models rely on a manually-driven parameter re-calibration mechanism, which takes an average of 12-15 weeks to complete model update verification, resulting in a process drift detection delay of more than 5 production batches. Especially in the large-scale mass production of 12-inch wafers, the lag in real-time optimization of process parameters will lead to a loss of overall batch yield of 2.3%-4.1% (data from the SEMI 2023 industry report).

[0005] The quality inspection system has structural defects: The current industry adopts a sampling plan with a full-process physical inspection coverage rate of less than 12%. The single-point detection cost of online metrology equipment is as high as $3.2 / measurement (data from VLSI Research), forcing enterprises to adopt a detection frequency compression strategy with increasing risks, resulting in the detection rate of abnormal process fluctuations dropping below 78%, significantly increasing the hidden quality cost in the mass production stage.

[0006] The existing technology improvement solutions have architectural defects: They adopt isolated data analysis modules or single-point prediction models, lacking a full-stack architecture covering data cleaning - feature engineering - multi-model fusion - closed-loop control. Specifically manifested as: 1. A multi-source heterogeneous data fusion mechanism for semiconductor manufacturing has not been constructed, resulting in a spatio-temporal alignment error of more than 30% in the training data set; 2. The collaborative optimization of models lacks process knowledge guidance, and the confidence level of prediction results in complex process scenarios is lower than 85%; 3. A dynamic coupling model of process parameters - equipment status - environmental variables has not been established, resulting in a closed-loop response delay of the production line control exceeding 8 hours, severely restricting the yield ramp-up speed of advanced processes. Summary of the Invention

[0007] The object of the present invention is to propose an intelligent prediction closed-loop control method, system and readable storage medium for semiconductor manufacturing, which can achieve high-precision prediction, real-time monitoring and automatic correction of process parameters in the semiconductor manufacturing process, and improve the prediction accuracy and reliability.

[0008] To achieve the above object, the present invention proposes an intelligent prediction closed-loop control method for semiconductor manufacturing, and the method includes the following steps: Obtain the real-time FDC data of the machine tool, input the FDC data into a preset virtual measurement model, and output the first Inline measurement parameter of the work in process, where the first Inline measurement parameter is the predicted value of the unmeasured Inline measurement parameter; Input the third Inline measurement parameter into a preset Inline-Wat prediction model, and output the predicted Wat value of the work in process. The third Inline measurement parameter is the sum of the first Inline measurement parameter and the second Inline measurement parameter, and the second Inline measurement parameter is the measured Inline measurement parameter; When it is judged that the predicted Wat value exceeds the abnormal control range, correct the Inline measurement parameters of the current process and the subsequent process of the product, so that the predicted Wat value output by the Inline-Wat prediction model after the corrected Inline measurement parameters does not exceed the abnormal control range.

[0009] Further, in the above semiconductor manufacturing intelligent prediction closed-loop control method, the steps of obtaining the real-time FDC data of the machine tool, inputting the FDC data into a preset virtual measurement model, and outputting the first Inline measurement parameter of the work in process include: Obtain the historical data of the machine tool, where the historical data includes the FDC data of the production equipment and the second Inline measurement parameter, and preprocess the historical data, where the second Inline measurement parameter is the measured Inline measurement parameter; Construct a virtual measurement model, and train and optimize the virtual measurement model according to the preprocessed historical data; Obtain the real-time FDC data of the machine tool, input the FDC data into the optimized virtual measurement model, and output the first Inline measurement parameter of the work in process.

[0010] Further, in the above semiconductor manufacturing intelligent prediction closed-loop control method, after the step of preprocessing the historical data, it further includes: Select the preprocessed historical data as the input of the virtual measurement model, calculate the prediction residual value R, residual mean M, and residual standard deviation S of the virtual measurement model, where the prediction residual value R is the difference between the real Inline measurement value and the first Inline measurement parameter; Judge whether the residual value R is within the interval [M - 3S, M + 3S]. If not, remove the data point; if so, retain the data point.

[0011] Further, in the above semiconductor manufacturing intelligent prediction closed-loop control method, the steps of constructing a virtual measurement model and training and optimizing the virtual measurement model according to the preprocessed historical data include: Construct a data set from the preprocessed historical data, where the data set includes a training set and a test set; Establish a virtual measurement model, generate the key hyperparameter combinations corresponding to the model on the training set, and perform cross-validation on each of the key hyperparameter combinations one by one to obtain the mean absolute percentage error score of the virtual measurement model on the training set, and use the key hyperparameter corresponding to the smallest mean absolute percentage error score as the optimal hyperparameter; Train the virtual measurement model on the training set with the optimal hyperparameter; Calculate the mean absolute percentage error score of the virtual measurement model on the test set, and use the virtual measurement model with the smallest mean absolute percentage error score on the test set as the optimal model.

[0012] Further, in the above semiconductor manufacturing intelligent prediction closed-loop control method, the step of performing cross-validation on each of the key hyperparameter combinations one by one includes: Divide the training set into K subsets, select K - 1 subsets to train the virtual measurement model, and use the remaining 1 subset to verify the virtual measurement model. Repeat the above process iteratively to obtain the mean absolute percentage error score of the virtual measurement model on the training set.

[0013] Further, in the above semiconductor manufacturing intelligent prediction closed-loop control method, the step of inputting the third Inline measurement parameter into a preset Inline-Wat prediction model to output the predicted Wat value of the work in process includes: Perform filtering processing on the third Inline measurement parameter; Calculate the correlation coefficient between the filtered third Inline measurement parameter and the Wat value; Determine the contribution weights corresponding to the third Inline measurement parameters at different levels, and calculate the comprehensive contribution weight of the third Inline measurement parameter to the Wat value according to the correlation coefficient and the contribution weights; Based on the comprehensive contribution weights, determine the optimal Inline measurement parameter set by the recursive feature method. Construct multiple prediction models according to the optimal Inline measurement parameter set, calculate the mean absolute percentage error scores of the multiple prediction models on the test set, and select the model with the smallest mean absolute percentage error score as the Inline-Wat prediction model; Input the filtered third Inline measurement parameter into the Inline-Wat prediction model to output the predicted Wat value of the work in process.

[0014] Further, in the above semiconductor manufacturing intelligent prediction closed-loop control method, the step of determining the optimal Inline measurement parameter set by the recursive feature method based on the comprehensive contribution weights specifically includes: Determine an initial parameter set from the third Inline measurement parameters based on the comprehensive contribution weights; Divide the initial parameter set into a training set and a test set, and construct an xgboost model on the training set to train the initial parameter set; Iteratively calculate the mean absolute percentage error scores of all parameters in the initial parameter set on the test set, and select the parameter set with the smallest mean absolute percentage error score as the optimal Inline measurement parameter set.

[0015] Further, in the above semiconductor manufacturing intelligent prediction closed-loop control method, the third Inline measurement parameters include: thickness parameters, dimension parameters, material parameters, and product parameters.

[0016] In addition, the present invention also provides a semiconductor manufacturing intelligent prediction closed-loop control system, including: A data acquisition unit for acquiring real-time FDC data of a machine tool; A first prediction unit for inputting the FDC data into a preset virtual measurement model and outputting a first Inline measurement parameter of a work in process, where the first Inline measurement parameter is a predicted value of an unmeasured Inline measurement parameter; A second prediction unit for inputting a third Inline measurement parameter into a preset Inline-Wat prediction model and outputting a predicted Wat value of a work in process, where the third Inline measurement parameter is the sum of the first Inline measurement parameter and the second Inline measurement parameter, and the second Inline measurement parameter is a measured Inline measurement parameter; A correction feedback unit for, when it is determined that the predicted Wat value exceeds an abnormal control range, correcting the Inline measurement parameters of the current process and the subsequent process of the product so that the predicted Wat value output by the Inline-Wat prediction model after the corrected Inline measurement parameters does not exceed the abnormal control range.

[0017] In addition, the present invention also provides a readable storage medium on which a computer program is stored, and the program is executed by a processor to implement the above semiconductor manufacturing intelligent prediction closed-loop control method.

[0018] The semiconductor manufacturing intelligent prediction closed-loop control method, system and readable storage medium of the present invention realize high-precision prediction, real-time monitoring and automatic correction of process parameters in the semiconductor manufacturing process through a multi-model collaborative optimization and closed-loop control architecture, improve the prediction accuracy and reliability, and maintain high prediction ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flowchart of the semiconductor manufacturing intelligent prediction closed-loop control method according to an embodiment of the present invention; Figure 2 is Figure 1 a specific process schematic diagram of step S1 in Figure 3 is Figure 1 a specific process schematic diagram of step S2 in Figure 4 is a structural schematic diagram of the semiconductor manufacturing intelligent prediction closed-loop control system according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] In this embodiment, the semiconductor manufacturing intelligent prediction closed-loop control method and system are taken as examples, and the present invention will be described in detail below with reference to specific embodiments and drawings.

[0021] The present invention. A semiconductor manufacturing intelligent prediction closed-loop control method provided by an embodiment of the present invention includes the following steps: obtaining real-time FDC (Fault Detection and Classification) data of a machine tool, inputting the FDC data into a preset virtual metrology model, and outputting a first Inline metrology parameter of a work in process, where the first Inline metrology parameter is a predicted value of an unmeasured Inline metrology parameter; inputting a third Inline metrology parameter into a preset Inline-Wat prediction model, and outputting a predicted Wat value of the work in process, where the third Inline metrology parameter is the sum of the first Inline metrology parameter and a second Inline metrology parameter, and the second Inline metrology parameter is a measured Inline metrology parameter; when it is determined that the predicted Wat value exceeds an abnormal control range, correcting the Inline metrology parameters of the current process and the subsequent process of the product so that the predicted Wat value output by the Inline-Wat prediction model after the corrected Inline metrology parameters do not exceed the abnormal control range. Through multi-model collaborative optimization and a closed-loop control architecture, the present invention realizes high-precision prediction, real-time monitoring, and automatic correction of process parameters in the semiconductor manufacturing process. The model has real-time monitoring capabilities, can dynamically analyze production process data, and quickly feedback deviations or abnormalities; by processing complex relationships between variables, the prediction accuracy and reliability are improved; it supports retraining the model when the production process is improved, continuously optimizing and updating, and maintaining high-efficiency prediction capabilities.

[0022] Please refer to Figure 1 , a semiconductor manufacturing intelligent prediction closed-loop control method provided by an embodiment of the present invention, the method specifically includes the following steps: Step S1: Obtain real-time FDC data of a machine tool, input the FDC data into a preset virtual metrology model, and output a first Inline metrology parameter of a work in process; In specific implementation, Inline means that a certain device, process, or task is directly integrated into the production system and is executed in real time and controlled by the production system (such as a manufacturing execution system MES). In the Inline state, the device or process is part of the production line and directly processes the tasks assigned by the system.

[0023] Inline measurement refers to the monitoring and measurement carried out in real time during the wafer manufacturing process. These measurements are usually non-destructive and are mainly used to check the production progress and process parameters in real time at each process step of manufacturing. The main purpose of Inline measurement is to ensure that deviations in each process step can be detected at an early stage, thus avoiding the continued production of defective wafers. This approach helps to reduce waste and improve overall production efficiency. Inline monitoring usually uses various advanced measurement tools, such as optical microscopes, scanning electron microscopes (SEM), optical measurement equipment, and probe measurement equipment, etc., to monitor key parameters such as the wafer surface, interlayer alignment (overlay), etching depth, and film thickness.

[0024] The present invention first presets to construct a virtual measurement model, trains the virtual measurement model through historical data, then obtains real-time data such as machine real-time FDC and inputs it into the virtual measurement model, and outputs the first Inline measurement parameter of the work in process. The first Inline measurement parameter is a predicted Inline measurement parameter, that is, the predicted value of the unmeasured Inline measurement parameter. In this way, the prediction of key process parameters is realized, the physical measurement requirements are reduced, time and costs are saved, and potential damages caused by physical measurement are avoided.

[0025] Please refer to Figure 2 , the specific steps of step S1 include: Step S11: Obtain machine historical data, the historical data includes production equipment FDC data and the second Inline measurement parameter, and preprocess the historical data. The second Inline measurement parameter is the measured Inline measurement parameter; Step S12: Construct a virtual measurement model, and train and optimize the virtual measurement model according to the preprocessed historical data; Step S13: Obtain machine real-time FDC data, input the FDC data into the optimized virtual measurement model, and output the first Inline measurement parameter of the work in process.

[0026] Specifically, in step S11, due to the sampling measurement in Inline measurement, there is less measured data and more unmeasured data. For example, when a chip wafer passes through three process stations A, B, and C, it may only be sampled for Inline measurement at station A, and stations B and C are not Inline measured. Therefore, by constructing a virtual measurement model, collecting the second Inline measurement parameters (the measured Inline measurement parameters) and combining the corresponding FDC to train the virtual measurement model, learning the relationship between FDC and Inline measurement parameters, so for the unmeasured data, by inputting their FDC data into the virtual measurement model, the first Inline measurement parameters are output, that is, using the first Inline measurement parameters (the predicted values of Inline measurement parameters) to supplement the Inline measurement parameter samples and make up for the small amount of data caused by the missing Inline measurement parameter samples.

[0027] The steps of preprocessing the historical data include: Judging whether each data point in the historical data exceeds the preset target range. If so, removing the data point; if not, retaining the data point.

[0028] Among them, a specification range (i.e., the preset target range, parameter spec) is set for each Inline measurement parameter, and the out-of-specification data is filtered out through the target range control. For example, the preset target range of Inline A parameter is [2, 10]. For a data point with an Inline A parameter value of 100, then this data point will be filtered out.

[0029] In step S12, a virtual measurement model from FDC to predicted Inline measurement is constructed. By selecting FDC parameters and the second Inline measurement parameters as the model inputs, the first Inline measurement parameters of the work-in-progress are output, which can be used to solve the problem of too little data volume caused by sampling measurement in Inline data during subsequent Inline-Wat prediction. It is not only possible to select the xgboost model, but other models, such as random forest, can also be used, as long as it is a model that can be used for regression prediction tasks.

[0030] After obtaining the first Inline measurement parameters of the work-in-progress, abnormal points will also be filtered, that is, selecting the preprocessed historical data as the model input, calculating the model prediction residual value R (residual) = true Inline measurement value - predicted Inline measurement value, the residual mean M, and the residual standard deviation S, and judging whether the residual value is within the interval [M - 3S, M + 3S]. If so, retaining it; if not, filtering out the data point.

[0031] That is, after the step of preprocessing the historical data, the following steps are further included: Select the preprocessed historical data as the input of the virtual measurement model, and calculate the prediction residual value R, the residual mean value M, and the residual standard deviation S of the virtual measurement model. The prediction residual value R is the difference between the true Inline measurement value and the first Inline measurement parameter; Judge whether the residual value R is within the interval [M - 3S, M + 3S]. If not, remove the data point; if so, retain the data point.

[0032] In this embodiment, the virtual measurement model is an xgboost model. The virtual measurement model can also be other models such as random forest, as long as it can be used for regression prediction tasks.

[0033] Specifically, in step S12, the following models can be selected for the virtual measurement model: xgboost model, DNN model, and random forest model (Random Forest).

[0034] Among them, the xgboost model is specifically: Model principle: Xgboost is an ensemble learning algorithm based on gradient boosting. It sequentially constructs multiple decision trees, and each tree focuses on correcting the prediction residuals of the previous tree.

[0035] Its core optimizations include: Loss function + regularization: The objective function includes a loss function (such as mean squared error, cross entropy) and a regularization term (L1 / L2) to control the model complexity.

[0036] Second-order Taylor expansion: Utilize the first and second derivatives of the loss function to more accurately update the model parameters.

[0037] Parallelization and cache optimization: Perform parallel computing during feature selection and accelerate the search for split points through pre-sorting.

[0038] Step S12 specifically includes: Select the xgboost model as the virtual measurement model; Initialization process: Use the initial prediction value (such as the mean value) as the baseline model.

[0039] Iterative tree building: Calculate the residuals of the current model (the negative gradient of the loss function); Generate a decision tree based on the residuals, and select the feature and threshold that maximize the gain of the objective function during each split; Accelerate the construction of the virtual measurement model by using a greedy algorithm or an approximation algorithm (such as histogram optimization); Model training update: The prediction results of the new tree are weighted (learning rate control) and accumulated into the current model; When the virtual measurement model reaches the preset number of trees (n_estimators) or the residual convergence value, stop the model training update.

[0040] The key hyperparameters of the virtual measurement model include: learning rate (learning_rate), maximum depth of the tree (max_depth), subsample ratio (subsample), and number of trees (n_estimators).

[0041] In step S12, when the virtual measurement model selects a DNN model, the DNN model is specifically: DNN (Deep Neural Network) Model principle: DNN is a neural network composed of multiple hidden layers, stacked through non-linear activation functions (such as ReLU, Sigmoid), to learn high-order abstract features of the input data. Its core includes: Forward propagation: The input data is calculated layer by layer to obtain the predicted output.

[0042] Backward propagation: Calculate the gradient according to the loss function (such as cross-entropy, mean squared error), and update the weights layer by layer.

[0043] Optimization algorithm: Such as Adam, SGD, dynamically adjust the learning rate.

[0044] The specific steps of S12 include: Select the DNN model as the virtual measurement model; Data preprocessing: Standardize / normalize the input, and divide the training set and validation set; Forward propagation: The input data is calculated through each layer (linear transformation + activation function) to obtain the output; Calculate the loss: Compare the difference between the predicted value and the true value (such as cross-entropy loss); Backward propagation: Calculate the gradient of the loss with respect to the weights of each layer using the chain rule; The optimizer updates the weights according to the gradient (such as gradient descent: w = w - η * ∇loss).

[0045] Regularization: Use Dropout, L2 regularization, or early stopping (Early Stopping) to prevent overfitting.

[0046] Iterative optimization: Repeat forward-backward propagation until convergence.

[0047] The key hyperparameters of the virtual measurement model include: learning rate, batch size (batch_size), number of hidden layers and neurons, and activation function.

[0048] In step S12, when the random forest model is selected as the virtual measurement model, the random forest model is specifically: Model principle: Random forest is an ensemble method based on Bagging, which constructs multiple decision trees and aggregates the results (classification voting, regression averaging). Its core randomness is reflected in: Sample randomness: Each tree is generated based on a training set obtained by bootstrap sampling.

[0049] Feature randomness: When splitting nodes, the optimal feature is selected from a random subset.

[0050] Step S12 specifically includes: Select the random forest model as the virtual measurement model; Bootstrap sampling: Samples are drawn from the original data with replacement to generate multiple groups of training subsets; Parallel tree construction: Independently train a decision tree for each subset; (When splitting nodes, only search for the best split point from a randomly selected subset of features) Aggregate results: The prediction results of all trees are obtained through voting (classification) or averaging (regression) to get the final output.

[0051] The key hyperparameters of the virtual measurement model include: number of trees (n_estimators), maximum depth of each tree (max_depth), size of the feature subset (max_features).

[0052] Step S12 specifically includes: Construct a data set from the preprocessed historical data, and the data set includes a training set and a test set; In this implementation, the proportion of the training set is 0.8, and the proportion of the test set is 0.2; Establish a virtual measurement model, generate key hyperparameter combinations corresponding to the model on the training set, and perform cross-validation on each key hyperparameter combination one by one to obtain the mean absolute percentage error (MAPE) score of the virtual measurement model on the training set. The key hyperparameter corresponding to the minimum mean absolute percentage error score is the optimal hyperparameter; Train the virtual measurement model on the training set with the optimal hyperparameters; Calculate the mean absolute percentage error (MAPE) score of the virtual measurement model on the test set, and the virtual measurement model with the minimum mean absolute percentage error score on the test set is the optimal model.

[0053] Specifically, the key hyperparameter combinations vary according to different models. For example, the hyperparameter combinations for a random forest are: 'n_estimators': [50, 100, 200],'max_depth': [5, 10, 20],'max_features': [0.5, 0.8]. Traverse this key hyperparameter combination and verify the model one by one.

[0054] The step of performing cross-validation on the key hyperparameter combinations one by one includes: Divide the training set into K subsets (e.g., K = 5), select K - 1 subsets to train the virtual measurement model, and use the remaining 1 subset to verify the virtual measurement model. Traverse and repeat the above process to obtain the mean absolute percentage error score of the virtual measurement model on the training set.

[0055] Step S2: Input the third Inline measurement parameter into a preset Inline-Wat prediction model, and output the predicted Wat value of the work in process. The third Inline measurement parameter is the sum of the first Inline measurement parameter and the second Inline measurement parameter; In specific implementation, first process the Inline measurement parameter and the predicted Inline measurement parameter through an AI data filtering layer. Among them, the Inline measurement parameter is the measured real data, and the predicted Inline measurement parameter is the predicted data of the virtual measurement model, which is used to supplement the data missing caused by the sampling measurement of Inline, and expand the sample size for subsequent Inline-Wat modeling. Secondly, extract multi-dimensional parameters through a correlation analysis layer and input them into a causality analysis layer to optimize the weights; finally, output the Wat prediction result through a recursive modeling layer to obtain the predicted Wat value of the work in process.

[0056] Please refer to Figure 3 , and the specific steps of step S2 include: Step S21: Filter the third Inline measurement parameter; Step S22: Calculate the correlation coefficient between the filtered third Inline measurement parameter and the Wat value; Step S23: Determine the contribution weights corresponding to the third Inline measurement parameters at different levels, and calculate the comprehensive contribution weight of the third Inline measurement parameter to the Wat value according to the correlation coefficient and the contribution weights; Step S24: Determine the optimal Inline measurement parameter set through a recursive feature method based on the comprehensive contribution weights. Construct multiple prediction models according to the optimal Inline measurement parameter set, calculate the mean absolute percentage error (MAPE) scores of the multiple prediction models on the test set, and select the model with the smallest mean absolute percentage error score as the Inline-Wat prediction model; Step S25: Input the filtered third Inline measurement parameter into the Inline-Wat prediction model, and output the predicted Wat value of the work in process.

[0057] Specifically, in step S21, based on the Inline-Wat correlation weight analysis and combined with the batch information (LOT ID), consistency processing is performed on the data of different batches. An anomaly filtering model (such as xgboost) is constructed to filter out the abnormal data points whose residual values are outside the 3-sigma range, providing a high-quality data basis for subsequent model training.

[0058] The specific steps of step S21 include: Select the third Inline measurement parameter with a non-zero correlation weight and the batch information to construct an anomaly filtering model; Use the third Inline measurement parameter as the input of the anomaly filtering model, calculate the predicted residual value R', residual mean M', and residual standard deviation S' of the anomaly filtering model, and determine whether the residual value R' is within the interval [M' - 3S', M' + 3S']. If not, remove the data point; the predicted residual value R' is the difference between the true Wat value and the predicted Wat value.

[0059] That is, select the third Inline measurement parameter and the batch information as the model input, calculate the model predicted residual value R' (residual) = true Wat value - predicted Wat value, residual mean M' (residual_mean), and residual standard deviation S' (sigma), and determine whether the residual value R' falls within [M' - 3S', M' + 3S']. If not, filter out the data point; if so, retain the data point.

[0060] Specifically, in step S22, the correlation between the third Inline measurement parameter and the corresponding Wat value is also analyzed through a correlation analysis layer, and the correlation coefficient between the third Inline measurement parameter and the corresponding Wat value is calculated.

[0061] The correlation coefficient is a statistical index that measures the strength and direction of the linear relationship between two variables. The correlation coefficient is a statistical index that measures the strength and direction of the linear relationship between two variables.

[0062] The formula for the correlation coefficient is as follows: X represents the third Inline measurement parameter; Y represents the Wat value corresponding to X, and X - represents the average value of the third Inline measurement parameter, and Y - represents the average value of the Wat value corresponding to X.

[0063] The value range of the correlation coefficient is between -1 and 1. The closer its absolute value is to 1, the stronger the linear relationship between the two variables; the closer the absolute value is to 0, the weaker the linear relationship between the two variables. When the correlation coefficient is positive, it indicates that the two variables are positively correlated; when the correlation coefficient is negative, it indicates that the two variables are negatively correlated.

[0064] The third Inline measurement parameter includes multi-dimensional inline process parameters. For example: thickness parameters: dielectric thickness, metal thickness; dimension parameters: lithography CD, after-etch CD; material parameters: contact metal liner material, metal type; product parameters: Product ID, Suffix No, Site No, Key Recipe, etc.

[0065] For example, by analyzing the correlation between the third Inline measurement parameter and the corresponding Wat value, it is obtained that the correlation coefficient between the after-etch CD and the Wat value is -0.76; this indicates that the smaller the size after etching, the higher the resistance, which may be because the line width becomes narrower, resulting in an increase in resistance; at the same time, the correlation coefficient between the metal thickness and the Wat value is -0.69, also showing a negative correlation. Classify and integrate the above parameters to construct a multi-dimensional parameter system, and mark the above parameters as key input variables for the next step of modeling.

[0066] Specifically, in step S23, first, analyze the correlation between the third Inline measurement parameters. By analyzing the correlation between intra-layer and cross-layer parameters, use statistical analysis and machine learning algorithms to filter out irrelevant parameters without causality, that is, judge the strength of the correlation between the Inline measurement parameter and the Wat value according to the correlation coefficient, and select the Inline measurement parameter with a strong correlation. For example, select the Inline measurement parameter with a correlation coefficient above 0.3.

[0067] Secondly, determine the contribution weights corresponding to the third Inline measurement parameters at different levels. Refer to Table 1. Six Inline measurement parameters with strong correlations are selected in Table 1. Based on the film layer of the chip structure, the contribution weight corresponding to the Inline measurement parameter in the current layer Metal2 is set to 1, while the adjacent layers Metal1, Via1, and Contact are given decreasing weights in sequence according to their physical distances from the current layer Metal2. For cross-layers, the weights are attenuated step by step according to the distances. In this embodiment, the attenuation rate is set to decrease the weight by 0.1 for each additional layer of distance, that is, the contribution weights corresponding to the third Inline measurement parameters at different levels are determined to be 1, 1, 0.9, 0.9, 0.9, and 0.8 in sequence.

[0068] Finally, calculate the comprehensive contribution weight of the third Inline measurement parameter to the Wat value according to the correlation coefficient and the contribution weight. For example, in Table 1, the comprehensive contribution weight of the metal thickness parameter (Thickness) of the current layer Metal2 to the Wat value is 0.75 × 1.0 = 0.75. By analogy, calculate the comprehensive contribution weights of all 6 Inline measurement parameters to the Wat value.

[0069] Table 1 Hierarchy Parameter Source layer Correlation coefficient Layer spacing Contribution weight Comprehensive contribution weight = coefficient × weight Metal2 Thickness Metal2 0.75 0 1 0.75 × 1.0 = 0.75 Metal2 CD Metal2 0.68 0 1 0.68 × 1.0 = 0.68 Metal1 Thickness Metal1 0.65 1 0.9 0.65 × 0.9 = 0.585 Metal1 CD Metal1 0.6 1 0.9 0.60 × 0.9 = 0.54 Via1 Thickness Via1 0.45 1 0.9 0.45 × 0.9 = 0.405 Contact Liner Mat. Contact 0.35 2 0.8 0.35 × 0.8 = 0.28 In this way, by adjusting the weight coefficient based on the causal relationship, output the contribution weights of each parameter in the third Inline measurement parameter to the Wat value, optimize the accuracy and interpretability of the model, and ensure that the prediction is based on the causal logic of actual production.

[0070] Specifically, step S24 uses the method of recursive feature selection to determine the optimal set of Inline measurement parameters with better prediction effects on Wat.

[0071] That is, the step of determining the optimal set of Inline measurement parameters by the recursive feature method based on the comprehensive contribution weight specifically includes: Determine the initial parameter set among the third Inline measurement parameters based on the comprehensive contribution weight; Divide the initial parameter set into a training set and a test set, and build an xgboost model on the training set to train the initial parameter set; Traverse and calculate the mean absolute percentage error (MAPE) scores of all parameters in the initial parameter set on the test set, and select the parameter set with the smallest mean absolute percentage error score as the optimal set of Inline measurement parameters.

[0072] Specifically, an initial parameter set A is determined from the third Inline measurement parameter based on the comprehensive contribution weight. The selected parameter set S = {}, best_mape is a sufficiently large value such as 100, and eps is a sufficiently small value such as 1e-4. The data set is divided into a training set and a test set, with the training set ratio being 0.8 and the test set ratio being 0.2.

[0073] Step a: Traverse the parameter set A - S once. The selected parameter is denoted as a, the best MAPE for this round is denoted as cur_best_mape, and the optimal parameter is best_parameter. Step b: Select the parameter set S + {a} and build an xgboost model on the training set. Step c: Calculate the MAPE score cur_mape of the test set. If cur_mape is less than cur_best_mape, then the value of cur_best_mape is cur_mape, and the value of best_parameter is a.

[0074] If cur_best_mape - best_mape > eps, add the parameter best_parameter to the parameter set S and go to Step a; otherwise, stop the loop.

[0075] After obtaining the optimal Inline measurement parameter set S, build multiple models: xgboost, DNN, RandomForest. Select the model with the minimum MAPE on the test set as the optimal model, i.e., the Inline-Wat prediction model, and input the predicted Inline measurement parameter into the Inline-Wat prediction model to output the predicted Wat value of the work-in-progress.

[0076] Step S3: Determine whether the predicted Wat value exceeds the abnormal control range. If so, perform Step S4; if not, do not process. In specific implementation, when predicting new data of the work-in-progress, the unmeasured Inline measurement parameters are supplemented with the set target values of the process parameters. The predicted Wat value of the work-in-progress is synchronously transmitted to the work-in-progress Wat real-time prediction and monitoring system. To detect abnormalities in a timely manner, the present invention sets an abnormal control range, including abnormal determination rules such as OOC (Out Of Control) and OOS (Out of Spec). If the predicted Wat value is within the abnormal control range, no operation is performed; if the predicted Wat value exceeds the abnormal control range, an alarm for abnormal products is triggered in a timely manner to link with the subsequent analysis system.

[0077] Step S4: Modify the Inline measurement parameters of the current process and subsequent processes of the product so that the predicted Wat value obtained by outputting the modified Inline measurement parameters through the Inline-Wat prediction model does not exceed the abnormal control range; In specific implementation, new data of the chip wafer is predicted through the Inline-Wat prediction model. For unmeasured Inline parameters, the set target value of the process parameters is used for supplementation. When it is determined that the predicted Wat value exceeds the abnormal control range (OOC or OOS occurs), that is, when an abnormal alarm is triggered, the Device APC calculation platform analyzes the association between Inline and Wat and generates a correction plan; the Process APC system executes process parameter adjustment, that is, for the front-end process of the wafer and the measured or predicted Inline parameters, the system can no longer adjust and correct, and only the Inline measurement parameters (unpredicted Inline measurement parameters) of the current process and subsequent processes of the product are corrected so that the Wat value re-predicted by the modified Inline measurement parameters through the Inline-Wat prediction model again does not exceed the abnormal control range. In this way, a closed-loop control link of "data input - prediction analysis - abnormal handling - process correction", forward prediction - reverse adjustment - post-adjustment prediction is formed.

[0078] Specifically, considering the case of multiple inputs and a single output, the input is x = (x1,..., xn), and the output of the Inline-Wat model f(x) is y'. When y' exceeds the set range of y (SPEC): [y low , y high , y low , y high are the lower and upper limits of y respectively. At this time, correcting y by adjusting x is equivalent to solving an optimization problem: ; ; ; When setting the abnormal control range of the predicted Wat value y, that is, [y low , y high , at this time, the control range of x can be obtained by solving the above formula as:

[0079] In this way, only the Inline measurement parameters (unpredicted Inline measurement parameters) of the current process and subsequent processes of the product need to be corrected, and the Inline measurement parameters of the current process and subsequent processes of the product are adjusted to [Xmin , X max Within this range, it can ensure that the Wat value re-predicted by the Inline-Wat prediction model for this Inline measurement parameter does not exceed the abnormal control range.

[0080] Please also refer to Figure 4 , the present invention also provides a semiconductor manufacturing intelligent prediction closed-loop control system for implementing the above semiconductor manufacturing intelligent prediction closed-loop control method. The system includes: A data acquisition unit 10 for acquiring real-time FDC data of the machine. A first prediction unit 20 for inputting the FDC data into a preset virtual measurement model and outputting a first Inline measurement parameter of the work in process. The first Inline measurement parameter is a predicted value of an unmeasured Inline measurement parameter. A second prediction unit 30 for inputting a third Inline measurement parameter into a preset Inline-Wat prediction model and outputting a predicted Wat value of the work in process. The third Inline measurement parameter is the sum of the first Inline measurement parameter and the second Inline measurement parameter, and the second Inline measurement parameter is a measured Inline measurement parameter. A correction feedback unit 40 for, when it is determined that the predicted Wat value exceeds the abnormal control range, correcting the Inline measurement parameters of the current process and the subsequent process of the product, so that the predicted Wat value output by the Inline-Wat prediction model for the corrected Inline measurement parameter does not exceed the abnormal control range.

[0081] In addition, the present invention also provides a readable storage medium, on which a computer program is stored. The program is executed by a processor to implement the semiconductor manufacturing intelligent prediction closed-loop control method as described above.

[0082] Compared with the prior art, the present invention has the following beneficial effects: 1. High-precision prediction: Through the multi-layer architecture design of the Wat prediction model and the processing ability of the virtual measurement model for complex relationships, the accuracy of electrical property prediction and measurement data prediction is improved.

[0083] 2. Real-time closed-loop control: The Smart APC closed-loop system realizes real-time response from data prediction, abnormal monitoring to process correction, shortens the problem handling cycle, and improves production efficiency.

[0084] 3. Cost optimization and adaptability improvement: The virtual measurement model reduces the need for physical measurement, reduces costs; the model supports dynamic training and updating to meet the requirements of the rapid iteration of semiconductor processes.

[0085] In summary, the intelligent prediction closed-loop control method, system, and readable storage medium of the present invention achieve high-precision prediction, real-time monitoring, and automatic correction of process parameters in the semiconductor manufacturing process through multi-model collaborative optimization and closed-loop control architecture, improving the prediction accuracy and reliability and maintaining the high-efficiency prediction ability.

[0086] The description and application of the present invention herein are illustrative and are not intended to limit the scope of the present invention to the above embodiments. Modifications and changes to the embodiments disclosed herein are possible, and substitutions and various equivalent components of the embodiments are known to those of ordinary skill in the art. It should be clear to those skilled in the art that the present invention can be implemented in other forms, structures, arrangements, proportions, and with other components, materials, and parts without departing from the spirit or essential characteristics of the present invention. Other modifications and changes can be made to the embodiments disclosed herein without departing from the scope and spirit of the present invention.

Claims

1. A semiconductor manufacturing intelligent prediction closed-loop control method, characterized in that The method includes the following steps: Obtain the real-time FDC data of the machine tool, input the FDC data into a preset virtual measurement model, and output the first Inline measurement parameter of the work in process. The first Inline measurement parameter is the predicted value of the unmeasured Inline measurement parameter. Input the third Inline measurement parameter into a preset Inline-Wat prediction model, and output the predicted Wat value of the work in process. The third Inline measurement parameter is the sum of the first Inline measurement parameter and the second Inline measurement parameter, and the second Inline measurement parameter is the measured Inline measurement parameter. When it is determined that the predicted Wat value exceeds the abnormal control range, correct the Inline measurement parameters of the current process and the subsequent process of the product, so that the predicted Wat value output by the Inline-Wat prediction model after the corrected Inline measurement parameters does not exceed the abnormal control range.

2. The semiconductor manufacturing intelligent prediction closed-loop control method according to claim 1, wherein The step of obtaining the real-time FDC data of the machine tool, inputting the FDC data into a preset virtual measurement model, and outputting the first Inline measurement parameter of the work in process includes: Obtain the historical data of the machine tool. The historical data includes the FDC data of the production equipment and the second Inline measurement parameter, and preprocess the historical data. The second Inline measurement parameter is the measured Inline measurement parameter. Construct a virtual measurement model, and train and optimize the virtual measurement model according to the preprocessed historical data. Obtain the real-time FDC data of the machine tool, input the FDC data into the optimized virtual measurement model, and output the first Inline measurement parameter of the work in process.

3. The semiconductor manufacturing intelligent prediction closed-loop control method according to claim 2, wherein After the step of preprocessing the historical data, it further includes: Select the preprocessed historical data as the input of the virtual measurement model, calculate the prediction residual value R, residual mean value M, and residual standard deviation S of the virtual measurement model. The prediction residual value R is the difference between the real Inline measurement value and the first Inline measurement parameter. Judge whether the residual value R is within the interval [M - 3S, M + 3S]. If not, remove the data point; if so, retain the data point.

4. The semiconductor manufacturing intelligent prediction closed-loop control method according to claim 2, wherein The step of constructing a virtual measurement model and training and optimizing the virtual measurement model according to the preprocessed historical data includes: Construct a data set from the preprocessed historical data. The data set includes a training set and a test set. Establish a virtual measurement model, generate the key hyperparameter combinations corresponding to the model on the training set, and perform cross-validation on each of the key hyperparameter combinations one by one to obtain the mean absolute percentage error score of the virtual measurement model on the training set. Take the key hyperparameter corresponding to the minimum mean absolute percentage error score as the optimal hyperparameter. Train the virtual measurement model on the training set with the optimal hyperparameter. Calculate the mean absolute percentage error score of the virtual measurement model on the test set, and take the virtual measurement model with the minimum mean absolute percentage error score on the test set as the optimal model.

5. The semiconductor manufacturing intelligent prediction closed-loop control method according to claim 4, wherein The step of performing cross-validation on each of the key hyperparameter combinations one by one includes: Dividing the training set into K subsets, selecting K - 1 subsets to train the virtual measurement model, and using the remaining 1 subset to validate the virtual measurement model. Repeat the above process iteratively to obtain the mean absolute percentage error score of the virtual measurement model on the training set.

6. The semiconductor manufacturing intelligent prediction closed-loop control method according to claim 1, characterized in that The step of inputting the third Inline measurement parameter into a preset Inline-Wat prediction model and outputting the predicted Wat value of the work in process includes : Performing filtering processing on the third Inline measurement parameter; Calculating the correlation coefficient between the filtered third Inline measurement parameter and the Wat value; Determining the contribution weights corresponding to the third Inline measurement parameters at different levels, and calculating the comprehensive contribution weight of the third Inline measurement parameter to the Wat value according to the correlation coefficient and the contribution weights; Based on the comprehensive contribution weights, determining the optimal set of Inline measurement parameters by means of recursive feature selection. Constructing multiple prediction models according to the optimal set of Inline measurement parameters, and calculating the mean absolute percentage error score of the multiple prediction models on the test set. Selecting the model with the smallest mean absolute percentage error score as the Inline-Wat prediction model; Inputting the filtered third Inline measurement parameter into the Inline-Wat prediction model and outputting the predicted Wat value of the work in process.

7. The semiconductor manufacturing intelligent prediction closed-loop control method according to claim 6, wherein, The step of determining the optimal set of Inline measurement parameters by means of recursive feature selection based on the comprehensive contribution weights specifically includes: Determining an initial parameter set from the third Inline measurement parameters based on the comprehensive contribution weights; Dividing the initial parameter set into a training set and a test set, and constructing an xgboost model on the training set to train the initial parameter set; Iteratively calculating the mean absolute percentage error score of all parameters in the initial parameter set on the test set, and selecting the parameter set with the smallest mean absolute percentage error score as the optimal set of Inline measurement parameters.

8. The semiconductor manufacturing intelligent prediction closed-loop control method according to claim 1, characterized in that The third Inline measurement parameters include: thickness parameters, dimension parameters, material parameters, and product parameters.

9. A semiconductor manufacturing intelligent prediction closed-loop control system for implementing the semiconductor manufacturing intelligent prediction closed-loop control method according to any one of claims 1-8, characterized in that, The system includes: A data acquisition unit for acquiring real-time FDC data of the machine; A first prediction unit for inputting the FDC data into a preset virtual measurement model and outputting the first Inline measurement parameter of the work in process, where the first Inline measurement parameter is the predicted value of the unmeasured Inline measurement parameter; A second prediction unit for inputting the third Inline measurement parameter into a preset Inline-Wat prediction model and outputting the predicted Wat value of the work in process, where the third Inline measurement parameter is the sum of the first Inline measurement parameter and the second Inline measurement parameter and, the second Inline measurement parameter is the measured Inline measurement parameter; A correction feedback unit, which is configured to correct the inline measurement parameters of the current process and subsequent processes of the product when it is determined that the predicted Wat value exceeds the abnormal control range, so that the predicted Wat value obtained by outputting the corrected inline measurement parameters through the Inline-Wat prediction model does not exceed the abnormal control range.

10. A readable storage medium, on which a computer program is stored, and the program is executed by a processor to implement the semiconductor manufacturing intelligent prediction closed-loop control method according to any one of claims 1-8.

10. A readable storage medium, on which a computer program is stored, and the program is executed by a processor to implement the semiconductor manufacturing intelligent prediction closed-loop control method according to any one of claims 1-8.

Images