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

Abstract

The present invention discloses a method, system, and readable storage medium for intelligent predictive closed-loop control of semiconductor manufacturing. The method includes 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 the work-in-progress; inputting a third inline measurement parameter into a preset inline-watt prediction model, and outputting a predicted watt value of the work-in-progress; and when it is determined that the predicted watt value exceeds an abnormal control range, correcting the inline measurement parameters of the current and subsequent processes of the work-in-progress. The method, system, and readable storage medium for intelligent predictive closed-loop control of semiconductor manufacturing of the present invention achieve high-precision prediction, real-time monitoring, and automated correction of process parameters of the semiconductor manufacturing process, improving prediction accuracy and reliability, and maintaining efficient prediction capabilities.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor manufacturing technology, and in particular to a semiconductor manufacturing intelligent predictive closed-loop control method, system, and readable storage medium. Background Art

[0002] In the field of advanced semiconductor manufacturing, as chip feature sizes evolve toward 3nm nodes and below, significant multi-physics coupling effects are emerging in the collaborative manufacturing process. Current production lines integrate over 500 process steps, and the high-order nonlinear interactions between process parameters are growing exponentially, posing unprecedented challenges to the precise control of the manufacturing process. Inline Watt (Wafer Acceptable Test) is a crucial concept in semiconductor manufacturing. It refers to Watt testing performed at a certain stage in the wafer manufacturing process (such as the inter-metal stage). The purpose of this test is to check the electrical performance parameters of the wafer in real time during the process, monitor wafer quality during production, and ensure that the electrical characteristics of each process step meet design requirements, thereby ensuring the yield of the finished product.

[0003] Traditional process modeling methods based on expert experience employ linear modeling methods based on simplified rules, exposing systemic flaws in addressing the following core issues: Inadequate representation of complex nonlinear features: Key process indicators (such as dielectric deposition thickness and photolithography critical dimensions (CD)) exist in strongly coupled nonlinear mapping relationships with Watt electrical parameters. Experimental data shows that traditional empirical models capture errors exceeding 35% in high-order interactions and time-varying features. This results in a root mean square error (RMSE) between process window predictions and measured values ​​that generally exceeds the industry's accuracy threshold, making it difficult to support sub-nanometer process control requirements.

[0004] Lack of dynamic process adaptation mechanisms: In an industry where iteration cycles for advanced materials like new high-k dielectrics and EUV photoresists have been shortened to 6-8 months, traditional models rely on manually driven parameter recalibration, requiring an average of 12-15 weeks to complete model update verification. This delays process drift detection by more than five production batches. Especially in large-scale 12-inch wafer production, delayed real-time process parameter optimization can result in a 2.3%-4.1% yield loss across the entire batch (according to the SEMI 2023 industry report).

[0005] The quality inspection system has structural flaws: the industry currently uses a sampling scheme with a full-process physical inspection coverage rate of less than 12%. The cost of single-point inspection with online metering equipment is as high as $3.2 / measurement (VLSI Research data). This forces companies to adopt a risk-increasing inspection frequency compression strategy, causing the detection rate of abnormal process fluctuations to drop below 78%, significantly increasing the hidden quality costs in the mass production stage.

[0006] Existing technical improvement solutions have architectural flaws: they use isolated data analysis modules or single-point prediction models, and lack a full-stack architecture covering data cleaning, feature engineering, multi-model fusion, and closed-loop control. Specifically,

[0007] 1. Failure to build a multi-source heterogeneous data fusion mechanism for semiconductor manufacturing resulted in over 30% spatiotemporal alignment errors in the training dataset.

[0008] 2. Model collaborative optimization lacks process knowledge guidance, and the confidence level of prediction results in complex process scenarios is less than 85%;

[0009] 3. Failure to establish a dynamic coupling model of process parameters, equipment status, and environmental variables resulted in a production line control closed-loop response delay of more than 8 hours, seriously restricting the yield ramp-up speed of advanced processes. Summary of the Invention

[0010] The purpose of the present invention is to propose an intelligent predictive 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 of the semiconductor manufacturing process, thereby improving prediction accuracy and reliability.

[0011] To achieve the above object, the present invention proposes a semiconductor manufacturing intelligent predictive closed-loop control method, the method comprising the following steps:

[0012] Acquire real-time FDC data of the machine, input the FDC data into a preset virtual measurement model, and output a first inline measurement parameter of the work-in-progress, where the first inline measurement parameter is a predicted value of an unmeasured inline measurement parameter;

[0013] Inputting a third inline measurement parameter into a preset inline-Wat prediction model to output a predicted Watt value of the work-in-progress, wherein 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;

[0014] When it is determined that the predicted Watt value exceeds the abnormal control range, the Inline measurement parameters of the current process and the subsequent process of the product are corrected so that the predicted Watt value obtained by the output of the Inline-Wat prediction model after the correction of the Inline measurement parameters does not exceed the abnormal control range.

[0015] Furthermore, in the above-mentioned semiconductor manufacturing intelligent predictive closed-loop control method, the step of obtaining real-time FDC data of the machine, inputting the FDC data into a preset virtual measurement model, and outputting the first inline measurement parameter of the work-in-progress includes:

[0016] Acquiring historical data of the machine, the historical data including FDC data of the production equipment and a second inline measurement parameter, and preprocessing the historical data, the second inline measurement parameter being a measured inline measurement parameter;

[0017] Constructing a virtual measurement model, and training and optimizing the virtual measurement model based on preprocessed historical data;

[0018] The real-time FDC data of the machine is obtained, the FDC data is input into the optimized virtual measurement model, and the first inline measurement parameters of the work-in-progress are obtained as output.

[0019] Furthermore, in the above-mentioned semiconductor manufacturing intelligent prediction closed-loop control method, after the step of preprocessing the historical data, the method further includes:

[0020] Select the preprocessed historical data as the input of the virtual measurement model, and calculate the predicted residual value R, residual mean M, and residual standard deviation S of the virtual measurement model, where the predicted residual value R is the difference between the true inline measurement value and the first inline measurement parameter;

[0021] Determine whether the residual value R is in the interval [M-3S, M+3S]. If not, remove the data point; if so, retain the data point.

[0022] Furthermore, in the above-mentioned semiconductor manufacturing intelligent prediction closed-loop control method, the step of constructing a virtual measurement model and training and optimizing the virtual measurement model based on preprocessed historical data includes:

[0023] Constructing a data set from the preprocessed historical data, wherein the data set includes a training set and a test set;

[0024] Establish a virtual measurement model, generate key hyperparameter combinations of the corresponding model on the training set, and cross-validate the key hyperparameter combinations one by one to obtain the mean absolute percentage error score of the virtual measurement model on the training set. The key hyperparameter corresponding to the minimum mean absolute percentage error score is determined as the optimal hyperparameter;

[0025] Training the virtual measurement model on a training set using the optimal hyperparameters;

[0026] The mean absolute percentage error score of the virtual measurement model on the test set is calculated, and the virtual measurement model with the smallest mean absolute percentage error score on the test set is selected as the optimal model.

[0027] Furthermore, in the above-mentioned semiconductor manufacturing intelligent predictive closed-loop control method, the step of cross-validating the key hyperparameter combinations one by one includes:

[0028] The training set is divided into K subsets, K-1 subsets are selected to train the virtual measurement model, and the remaining 1 subset is used to verify the virtual measurement model. The above process is repeated to obtain the mean absolute percentage error score of the virtual measurement model on the training set.

[0029] Furthermore, in the above-mentioned 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 a predicted Watt value of the work-in-progress includes:

[0030] Filter the third Inline measurement parameter;

[0031] Calculate the correlation coefficient between the filtered third inline measurement parameter and the Watt value;

[0032] Determine the contribution weights corresponding to the third inline measurement parameters at different levels, and calculate the comprehensive contribution weights of the third inline measurement parameters to the Watt value based on the correlation coefficients and the contribution weights;

[0033] Determining an optimal inline measurement parameter set based on the comprehensive contribution weights through a recursive feature approach, constructing multiple prediction models based on the optimal inline measurement parameter set, calculating mean absolute percentage error scores of the multiple prediction models on a test set, and selecting a model with the smallest mean absolute percentage error score as the inline-Wat prediction model;

[0034] The filtered third inline measurement parameter is input into the inline-Wat prediction model, and the predicted Watt value of the work-in-progress is output.

[0035] Furthermore, in the above-mentioned semiconductor manufacturing intelligent predictive closed-loop control method, the step of determining the optimal inline measurement parameter set based on the comprehensive contribution weight through a recursive feature method specifically includes:

[0036] Determining an initial parameter set from a third Inline measurement parameter based on the comprehensive contribution weight;

[0037] The initial parameter set is divided into a training set and a test set, and an xgboost model is constructed on the training set to train the initial parameter set;

[0038] The mean absolute percentage error scores of all parameters in the initial parameter set in the test set are calculated, and the parameter set with the smallest mean absolute percentage error score is selected as the optimal inline measurement parameter set.

[0039] Furthermore, in the above-mentioned semiconductor manufacturing intelligent predictive closed-loop control method, the third inline measurement parameters include: thickness parameters, size parameters, material parameters and product parameters.

[0040] In addition, the present invention also provides a semiconductor manufacturing intelligent prediction closed-loop control system, comprising:

[0041] Data acquisition unit, used to obtain real-time FDC data of the machine;

[0042] a first prediction unit, configured to input the FDC data into a preset virtual measurement model and output a first inline measurement parameter of the work-in-progress, where the first inline measurement parameter is a predicted value of an unmeasured inline measurement parameter;

[0043] a second prediction unit, configured to input a third inline measurement parameter into a preset inline-Wat prediction model and output a predicted Watt value of the work-in-process, wherein 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;

[0044] The correction feedback unit is used to correct the inline measurement parameters of the current process and the subsequent process of the product when it is determined that the predicted Watt value exceeds the abnormal control range, so that the predicted Watt value obtained by the output of the inline-Wat prediction model after the correction of the inline measurement parameters does not exceed the abnormal control range.

[0045] In addition, the present invention also provides a readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to implement the above-mentioned semiconductor manufacturing intelligent predictive closed-loop control method.

[0046] The semiconductor manufacturing intelligent predictive 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 of the semiconductor manufacturing process through multi-model collaborative optimization and closed-loop control architecture, thereby improving prediction accuracy and reliability and maintaining efficient prediction capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flow chart of a semiconductor manufacturing intelligent predictive closed-loop control method according to an embodiment of the present invention;

[0048] Figure 2 yes Figure 1 Specific flow diagram of step S1;

[0049] Figure 3 yes Figure 1 Specific flow diagram of step S2;

[0050] Figure 4 It is a structural diagram of a semiconductor manufacturing intelligent predictive closed-loop control system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0051] This embodiment takes the semiconductor manufacturing intelligent predictive closed-loop control method and system as an example, and the present invention will be described in detail below with reference to specific embodiments and drawings.

[0052] The present invention. An embodiment of the present invention provides an intelligent predictive closed-loop control method for semiconductor manufacturing, comprising the following steps: obtaining real-time FDC (Fault Detection and Classification) data of a machine, inputting the FDC data into a preset virtual measurement model, and outputting a first Inline measurement parameter of the work-in-progress, wherein the first Inline measurement parameter is a predicted value of an unmeasured Inline measurement parameter; inputting a third Inline measurement parameter into a preset Inline-Wat prediction model, and outputting a predicted Watt value of the work-in-progress, wherein 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; when it is determined that the predicted Watt value exceeds an abnormal control range, correcting the Inline measurement parameters of the current and subsequent processes of the product so that the predicted Watt value of the corrected Inline measurement parameter obtained by the Inline-Wat prediction model does not exceed the abnormal control range. The present invention achieves high-precision prediction, real-time monitoring and automatic correction of process parameters of the semiconductor manufacturing process through multi-model collaborative optimization and closed-loop control architecture. The model has real-time monitoring capabilities, can dynamically analyze production process data, and quickly feedback deviations or anomalies; by processing complex relationships between variables, it improves prediction accuracy and reliability; supports retraining of the model when the production process is improved, and continuously optimizes and updates to maintain efficient prediction capabilities.

[0053] See also Figure 1 An embodiment of the present invention provides a semiconductor manufacturing intelligent predictive closed-loop control method, the method specifically comprising the following steps:

[0054] Step S1: Acquire real-time FDC data of the machine, input the FDC data into a preset virtual measurement model, and output the first inline measurement parameters of the work-in-progress;

[0055] In practice, inline refers to the direct integration of a device, process, or task into a production system, allowing it to execute and be controlled in real time by that system (such as a Manufacturing Execution System (MES)). In this state, the device or process is part of the production line and directly handles tasks assigned by the system.

[0056] Inline metrology refers to real-time monitoring and measurement during the wafer manufacturing process. These measurements are usually non-destructive and are mainly used to check production progress and process parameters in real time during each manufacturing process step. The main purpose of inline metrology is to ensure that deviations in each process step can be detected early, thereby avoiding the continued production of unqualified wafers. This method helps to reduce waste and improve overall production efficiency. Inline monitoring usually uses various advanced measurement tools, such as optical microscopes, scanning electron microscopes (SEMs), optical metrology equipment, and probe measurement equipment to monitor key parameters such as the wafer surface, overlay, etching depth, and film thickness.

[0057] The present invention first constructs a virtual measurement model, trains it with historical data, then acquires real-time data, such as the machine's FDC, and inputs it into the virtual measurement model. The output is the first inline measurement parameter of the work-in-progress. This first inline measurement parameter is the predicted inline measurement parameter—the predicted value of the unmeasured inline measurement parameter. This enables prediction of key process parameters, reduces the need for physical measurement, saves time and costs, and avoids potential damage from physical measurement.

[0058] See also Figure 2 , the step S1 specifically includes:

[0059] Step S11: Acquire historical data of the machine, the historical data including FDC data of the production equipment and a second inline measurement parameter, and pre-process the historical data, the second inline measurement parameter being a measured inline measurement parameter;

[0060] Step S12: constructing a virtual measurement model, and training and optimizing the virtual measurement model based on the pre-processed historical data;

[0061] Step S13: Acquire the real-time FDC data of the machine, input the FDC data into the optimized virtual measurement model, and output the first inline measurement parameters of the work-in-progress.

[0062] Specifically, in step S11, due to sampling during inline measurement, there is relatively little measured data and relatively much unmeasured data. For example, a chip wafer passing through three process sites, A, B, and C, may only be sampled for inline measurement at site A, while sites B and C are both measured inline. Therefore, by constructing a virtual measurement model, collecting second inline measurement parameters (measured inline measurement parameters), and combining them with corresponding FDCs to train the virtual measurement model, the relationship between FDCs and inline measurement parameters is learned. Thus, for unmeasured data, their FDC data is input into the virtual measurement model, and the first inline measurement parameters are output. In other words, the first inline measurement parameters (predicted values ​​of the inline measurement parameters) are used to supplement the inline measurement parameter samples, compensating for the small amount of data caused by missing inline measurement parameter samples.

[0063] The step of preprocessing the historical data comprises:

[0064] Determine whether each data point in the historical data exceeds a preset target range. If so, remove the data point; if not, retain the data point.

[0065] Each inline measurement parameter has a specified range (i.e., a preset target range, parameter spec). This target range control is used to filter out data that exceeds the specified range. For example, if the preset target range for Inline A parameter is [2, 10], and a data point with a value of 100 is selected for Inline A, this data point will be filtered out.

[0066] In step S12, a virtual measurement model is constructed from FDC to predict inline measurement. By selecting the FDC parameters and the second inline measurement parameters as model inputs, the model outputs the first inline measurement parameters of the work-in-progress. This is used for subsequent inline-WAT predictions. This solves the problem of insufficient inline data due to sampling. The xgboost model is not the only choice; other models, such as random forests, can also be used, as long as they can be used for regression prediction tasks.

[0067] After obtaining the first inline measurement parameters of the work-in-process product, outliers are filtered out. That is, preprocessed historical data is selected as model input. The model prediction residual value R (residual) is calculated as: actual inline measurement value - predicted inline measurement value, residual mean M, residual standard deviation S. The residual value is then determined to be in the interval [M - 3S, M + 3S]. If so, the residual value is retained; if not, the data point is filtered out.

[0068] That is, after the step of preprocessing the historical data, the method further includes:

[0069] Select the preprocessed historical data as the input of the virtual measurement model, and calculate the predicted residual value R, residual mean M, and residual standard deviation S of the virtual measurement model, where the predicted residual value R is the difference between the true inline measurement value and the first inline measurement parameter;

[0070] Determine whether the residual value R is in the interval [M-3S, M+3S]. If not, remove the data point; if so, retain the data point.

[0071] In this embodiment, the virtual measurement model is an xgboost model. The virtual measurement model may also be other models such as random forest, as long as it can be used for regression prediction tasks.

[0072] Specifically, in step S12 , the virtual measurement model may be selected from the following models: xgboost model, DNN model, and random forest model.

[0073] The xgboost model is specifically:

[0074] Model principle: xgboost is an ensemble learning algorithm based on gradient boosting. It constructs multiple decision trees sequentially, with each tree focusing on correcting the prediction residual of the previous tree.

[0075] Its core optimizations include:

[0076] Loss function + regularization: The objective function includes the loss function (such as mean square error, cross entropy) and regularization terms (L1 / L2) to control the complexity of the model.

[0077] Second-order Taylor expansion: Utilizes the first and second-order derivatives of the loss function to update model parameters more accurately.

[0078] Parallelization and cache optimization: Parallel calculations are performed during feature selection, and split point search is accelerated by pre-sorting.

[0079] The step S12 specifically includes:

[0080] Select the xgboost model as the virtual measurement model;

[0081] Initialization processing: Use the initial predicted value (such as the mean) as the baseline model.

[0082] Iterative tree construction: Calculate the residual of the current model (negative gradient of the loss function);

[0083] Generate a decision tree based on the residual, and select the features and thresholds that maximize the objective function gain at each split;

[0084] Accelerate the tree construction of virtual measurement models through greedy algorithms or approximate algorithms (such as histogram optimization);

[0085] Model training update: The prediction results of the new tree are weighted (learning rate controlled) and accumulated to the current model;

[0086] When the virtual measurement model reaches a preset number of trees (n_estimators) or a residual convergence value, the model training and updating is stopped.

[0087] The key hyperparameters of the virtual measurement model include: learning rate (learning_rate), maximum tree depth (max_depth), subsample ratio (subsample), and number of trees (n_estimators).

[0088] In step S12, when the virtual measurement model selects a DNN model, the DNN model is specifically:

[0089] DNN (Deep Neural Network)

[0090] Model Principle: DNN is a neural network composed of multiple hidden layers, which learns high-level abstract features of input data by stacking nonlinear activation functions (such as ReLU and Sigmoid). Its core includes:

[0091] Forward propagation: The input data is calculated layer by layer to obtain the predicted output.

[0092] Backpropagation: Calculate the gradient based on the loss function (such as cross entropy, mean square error) and update the weights layer by layer.

[0093] Optimization algorithms: such as Adam and SGD, dynamically adjusting the learning rate.

[0094] The step S12 specifically includes:

[0095] Select the DNN model as the virtual measurement model;

[0096] Data preprocessing: standardize / normalize input, divide into training set and validation set;

[0097] Forward propagation: The input data is calculated through each layer (linear transformation + activation function) to obtain the output;

[0098] Calculate loss: compare the difference between the predicted value and the true value (such as cross entropy loss);

[0099] Backpropagation: The chain rule calculates the gradient of the loss with respect to each layer’s weights;

[0100] The optimizer updates the weights according to the gradient (e.g. gradient descent: w = w - η * ∇loss).

[0101] Regularization: Use Dropout, L2 regularization, or Early Stopping to prevent overfitting.

[0102] Iterative optimization: Repeat forward-backward propagation until convergence.

[0103] The key hyperparameters of the virtual measurement model include: learning rate, batch size (batch_size), number of hidden layers and neurons, and activation function.

[0104] In step S12, when the virtual measurement model selects the random forest model, the random forest model is specifically:

[0105] Model Principle: Random Forest is an ensemble method based on bagging, which builds multiple decision trees and aggregates the results (classification voting, regression averaging). Its core randomness is reflected in:

[0106] Random samples: Each tree is generated based on the training set using bootstrap sampling.

[0107] Feature Randomness: Select the best features from a random subset when splitting a node.

[0108] The step S12 specifically includes:

[0109] Select the random forest model as the virtual measurement model;

[0110] Bootstrapping: extract samples from the original data with replacement to generate multiple training subsets;

[0111] Parallel tree building: train a decision tree for each subset independently; (when splitting a node, only find the best split point from the randomly selected feature subset)

[0112] Aggregate results: The prediction results of all trees are voted (classification) or averaged (regression) to obtain the final output.

[0113] The key hyperparameters of the virtual measurement model include the number of trees (n_estimators), the maximum depth of each tree (max_depth), and the feature subset size (max_features).

[0114] The step S12 specifically includes:

[0115] The preprocessed historical data is used to construct a data set, which includes a training set and a test set. In this embodiment, the ratio of the training set is 0.8, and the ratio of the test set is 0.2.

[0116] Establish a virtual measurement model, generate key hyperparameter combinations of the corresponding model on the training set, and cross-validate the key hyperparameter combinations 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 MAPE score is determined as the optimal hyperparameter.

[0117] Training the virtual measurement model on a training set using the optimal hyperparameters;

[0118] Calculate the mean absolute percentage error of the virtual measurement model on the test set

[0119] The virtual measurement model with the smallest mean absolute percentage error (MAPE) score on the test set is the optimal model.

[0120] Specifically, the key hyperparameter combination varies depending on the model. For example, the hyperparameter combination for random forest is: 'n_estimators': [50, 100, 200], 'max_depth': [5, 10, 20], 'max_features': [0.5, 0.8]. This key hyperparameter combination is traversed and the model is verified one by one.

[0121] The step of cross-validating the key hyperparameter combinations one by one includes:

[0122] The training set is divided into K subsets (e.g., K=5), K-1 subsets are selected to train the virtual measurement model, and the remaining 1 subset is used to verify the virtual measurement model. The above process is repeated to obtain the mean absolute percentage error score of the virtual measurement model on the training set.

[0123] Step S2: Inputting a third inline measurement parameter into a preset inline-Wat prediction model to output a predicted Watt value of the work-in-progress, wherein the third inline measurement parameter is the sum of the first inline measurement parameter and the second inline measurement parameter;

[0124] In its implementation, the AI ​​data filtering layer first processes inline measurement parameters and predicted inline measurement parameters. The inline measurement parameters are real-world data, while the predicted inline measurement parameters are predicted data from a virtual measurement model. This is used to supplement data missing due to spot checks and expand the sample size for subsequent inline-Wat modeling. Next, the correlation analysis layer extracts multi-dimensional parameters, which are then input into the causal analysis layer for weight optimization. Finally, the recursive modeling layer outputs the Watt prediction results, resulting in the predicted Watt value for the work-in-process (WIP).

[0125] See also Figure 3 , the step S2 specifically includes:

[0126] Step S21: filtering the third inline measurement parameter;

[0127] Step S22: Calculate the correlation coefficient between the filtered third inline measurement parameter and the Watt value;

[0128] Step S23: Determine the contribution weights corresponding to the third inline measurement parameters at different levels, and calculate the comprehensive contribution weights of the third inline measurement parameters to the Watt value based on the correlation coefficient and the contribution weights;

[0129] Step S24: Determine the optimal inline measurement parameter set based on the comprehensive contribution weights using a recursive feature approach, construct multiple prediction models based on 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 MAPE score as the inline-Wat prediction model;

[0130] Step S25: Input the filtered third inline measurement parameter into the inline-Wat prediction model, and output the predicted Watt value of the work-in-progress.

[0131] Specifically, in step S21, based on Inline-Wat correlation weight analysis and combined with batch information (LOT ID), consistency processing is performed on different batches of data. An anomaly filtering model (such as xgboost) is constructed to filter out anomalous data points with residual values ​​outside the 3-sigma range, providing a high-quality data foundation for subsequent model training.

[0132] The step S21 specifically includes:

[0133] Select the third inline measurement parameter and batch information whose correlation weight is not 0 to build an anomaly filtering model;

[0134] The third inline measurement parameter is used as the input of the anomaly filtering model to calculate the predicted residual value R', residual mean M' and residual standard deviation S' of the anomaly filtering model. It is determined whether the residual value R' is in the interval [M'- 3 S', M'+ 3S']. If not, the data point is removed; the predicted residual value R' is the difference between the true Wat value and the predicted Wat value.

[0135] That is, the third inline measurement parameter and batch information are selected as model inputs, and the model prediction residual value R' (residual) = true Watt value - predicted Watt value, residual mean M' (residual_mean), residual standard deviation S' (sigma) are calculated. It is determined whether the residual value R' falls within [M'- 3 S', M'+ 3 S']. If not, the data point is filtered out; if so, the data point is retained.

[0136] Specifically, in step S22, the correlation between the third inline measurement parameter and the corresponding Watt value is analyzed by the correlation analysis layer, and the correlation coefficient between the third inline measurement parameter and the corresponding Watt value is calculated.

[0137] The correlation coefficient is a statistical indicator that measures the strength and direction of the linear relationship between two variables. The correlation coefficient is a statistical indicator that measures the strength and direction of the linear relationship between two variables.

[0138] The correlation coefficient formula is as follows:

[0139]

[0140] X represents the third Inline measurement parameter; Y represents the Watt value corresponding to X, - represents the average value of the third Inline measurement parameter, Y - Indicates the average value of Watt corresponding to X.

[0141] The correlation coefficient ranges from −1 to 1. The closer its absolute value is to 1, the stronger the linear relationship between the two variables; the closer its absolute value is to 0, the weaker the linear relationship between the two variables. A positive correlation coefficient indicates a positive correlation between the two variables; a negative correlation coefficient indicates a negative correlation between the two variables.

[0142] The third inline measurement parameter includes multi-dimensional inline process parameters, such as thickness parameters: dielectric thickness, metal thickness; dimension parameters: litho CD, after-etch CD; material parameters: contact metal liner material, metal type; product parameters: product ID, suffix number, site number, key recipe, etc.

[0143] For example, by analyzing the correlation between the third inline measurement parameter and the corresponding Watt value, we found that the correlation coefficient between post-etch CD and Watt value is -0.76, indicating that smaller post-etch dimensions increase resistance, likely due to increased resistance caused by narrower traces. Furthermore, the correlation coefficient between metal thickness and Watt value is -0.69, also showing a negative correlation. These parameters are categorized and integrated to construct a multidimensional parameter system, and these parameters are labeled as key input variables for the next step of modeling.

[0144] Specifically, in step S23, the correlation between the third inline measurement parameters is first analyzed. By analyzing the correlation between the parameters within and across layers, statistical analysis and machine learning algorithms are used to filter out non-causal correlated parameters. Specifically, the correlation coefficient is used to determine the strength of the correlation between the inline measurement parameter and the Watt value, and inline measurement parameters with strong correlation are selected. For example, inline measurement parameters with a correlation coefficient of 0.3 or above are selected.

[0145] Secondly, determine the contribution weights corresponding to the third Inline measurement parameters at different levels. Please refer to Table 1. In Table 1, select the six Inline measurement parameters with strong correlation. Based on the chip structure film layer, set the contribution weight corresponding to the Inline measurement parameter in the current layer Metal2 to 1, and the adjacent layers Metal1, Via1, and Contact are assigned decreasing weights in turn according to their physical distance from the current layer Metal2. The cross-layer attenuation is step-by-step according to the distance. In this embodiment, the attenuation rate is set to reduce 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, respectively.

[0146] Finally, the combined contribution weight of the third inline measurement parameter to the Watt value is calculated based on the correlation coefficient and contribution weight. For example, in Table 1, the combined contribution weight of the metal thickness parameter (Thickness) of the current layer Metal2 to the Watt value is 0.75 x 1.0 = 0.75. Similarly, the combined contribution weights of all six inline measurement parameters to the Watt value are calculated.

[0147] Table 1

[0148] Hierarchy parameter Source layer Correlation coefficient Layer distance 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

[0149] In this way, by adjusting the weight coefficient based on the causal relationship, the contribution weight of each parameter in the third Inline measurement parameter to the Wat value is output, the accuracy and interpretability of the model are optimized, and it is ensured that the prediction is based on the actual production causal logic.

[0150] Specifically, the step S24 uses a recursive feature selection method to determine the optimal Inline measurement parameter set for Watt prediction.

[0151] That is, the step of determining the optimal Inline measurement parameter set based on the comprehensive contribution weight through a recursive feature method specifically includes:

[0152] Determining an initial parameter set from a third Inline measurement parameter based on the comprehensive contribution weight;

[0153] The initial parameter set is divided into a training set and a test set, and an xgboost model is constructed on the training set to train the initial parameter set;

[0154] The mean absolute percentage error (MAPE) scores of all parameters in the initial parameter set in the test set are traversed and calculated, and the parameter set with the smallest mean absolute percentage error score is selected as the optimal inline measurement parameter set.

[0155] Specifically, an initial parameter set A is determined in 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, eps is a sufficiently small value such as 1e-4, and 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.

[0156] Step a: Traverse a round of parameter set AS, record the selected parameter as a, record the best MAPE of this round as cur_best_mape, and the optimal parameter as best_parameter;

[0157] Step b: Select the parameter set S + {a} and build the xgboost model on the training set;

[0158] Step c: Calculate the MAPE score cur_mape of the test set;

[0159] If cur_mape is less than cur_best_mape, the value of cur_best_mape is cur_mape and the value of best_parameter is a.

[0160] 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.

[0161] After obtaining the optimal set of inline measurement parameters S, multiple models are constructed: xgboost, DNN, and RandomForest. The model with the smallest MAPE on the test set is selected as the optimal model, namely the inline-Wat prediction model. The predicted inline measurement parameters are input into the inline-Wat prediction model, and the predicted Watt value of the work-in-progress is obtained as the output.

[0162] Step S3: Determine whether the predicted Watt value exceeds the abnormal control range. If so, proceed to step S4; if not, do not process;

[0163] In specific implementation, new data of the work-in-process is predicted, and the unmeasured inline measurement parameters are supplemented with the set target values ​​of the process parameters. The output predicted Watt value of the work-in-process is synchronously transmitted to the real-time Watt prediction monitoring system of the work-in-process. In order to detect abnormalities in time, the present invention sets an abnormal control range, including abnormality judgment rules such as OOC (Out Of Control) and OOS (Out of Spec). If the predicted Watt value is within the abnormal control range, no operation is performed; if the predicted Watt value exceeds the abnormal control range, an abnormal product alarm is triggered in time, and the subsequent analysis system is linked.

[0164] Step S4: Correcting the inline measurement parameters of the current process and the subsequent process of the product so that the predicted Watt value output by the inline-Wat prediction model after the correction of the inline measurement parameters does not exceed the abnormal control range;

[0165] In specific implementations, the Inline-Wat prediction model uses new wafer data for prediction. Unmeasured inline parameters are supplemented with the target values ​​of the process parameters. When the predicted Watt value is determined to be outside the abnormal control range (occurring out-of-control or out-of-specification), triggering an abnormality alarm, the Device APC computing platform analyzes the relationship between inline and Watt and generates a correction plan. The Process APC system then adjusts the process parameters. This means that the system cannot adjust or correct the measured or predicted inline parameters of the wafer's upstream process. Instead, it corrects the measured inline parameters (unpredicted inline parameters) of the current and downstream processes of the product, ensuring that the newly predicted Watt values ​​from the Inline-Wat prediction model remain within the abnormal control range. This creates a closed-loop control chain: "data input - prediction analysis - abnormality handling - process correction," and a forward prediction - backward adjustment - adjusted prediction.

[0166] Specifically, consider the case of multiple inputs and single output, the input is x = (x1, ..., xn), the output of the Inline-Wat model f (x) is y', when y' exceeds the y setting range (SPEC): [y low , y high ], y low 、y high are the lower and upper limits of y, respectively. At this point, modifying y by adjusting x is equivalent to solving an optimization problem:

[0167] ;

[0168] ;

[0169] ;

[0170]

[0171] When setting the abnormal control range of the predicted Wat value y, that is, [y low , y high ], then the control range of x can be obtained by the above formula:

[0172]

[0173] In this way, it is only necessary to modify the Inline measurement parameters (unpredicted Inline measurement parameters) of the current process and the subsequent process of the product, and adjust the Inline measurement parameters of the current process and the subsequent process of the product to [X min , Xmax ] interval, it can be ensured that the Watt value re-predicted by the Inline-Wat prediction model for the Inline measurement parameter does not exceed the abnormal control range.

[0174] See also Figure 4 The present invention also provides a semiconductor manufacturing intelligent prediction closed-loop control system for implementing the above-mentioned semiconductor manufacturing intelligent prediction closed-loop control method, the system comprising:

[0175] Data acquisition unit 10, used to obtain real-time FDC data of the machine;

[0176] A first prediction unit 20 is configured to input the FDC data into a preset virtual measurement model and output a first inline measurement parameter of the work-in-progress, where the first inline measurement parameter is a predicted value of an unmeasured inline measurement parameter;

[0177] a second prediction unit 30 configured to input a third inline measurement parameter into a preset inline-Wat prediction model and output a predicted Watt value of the work-in-process, wherein 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;

[0178] The correction feedback unit 40 is used to correct the inline measurement parameters of the current process and the subsequent process of the product when it is determined that the predicted Watt value exceeds the abnormal control range, so that the predicted Watt value obtained by the inline-Wat prediction model output of the corrected inline measurement parameters does not exceed the abnormal control range.

[0179] In addition, the present invention also provides a readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to implement the above-mentioned intelligent predictive closed-loop control method for semiconductor manufacturing.

[0180] Compared with the prior art, the present invention has the following beneficial effects:

[0181] 1. High-precision prediction: The multi-layer architecture design of the Watt prediction model, combined with the virtual measurement model's ability to handle complex relationships, improves the accuracy of electrical property prediction and measurement data prediction.

[0182] 2. Real-time closed-loop control: The Smart APC closed-loop system enables real-time response from data prediction, abnormality monitoring to process correction, shortening the problem-solving cycle and improving production efficiency.

[0183] 3. Cost optimization and improved adaptability: The virtual measurement model reduces the need for physical measurement and lowers costs. The model supports dynamic training and updates to meet the needs of rapid iteration of semiconductor processes.

[0184] In summary, the semiconductor manufacturing intelligent predictive 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 of the semiconductor manufacturing process through multi-model collaborative optimization and closed-loop control architecture, thereby improving prediction accuracy and reliability and maintaining efficient prediction capabilities.

[0185] The description and application of the present invention here are illustrative and are not intended to limit the scope of the present invention to the above-described embodiments. The deformation and changes of the embodiments disclosed here are possible, and the replacement of the embodiments and the various components that are equivalent are well 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 realized 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 deformation and changes can be made to the embodiments disclosed here without departing from the scope and spirit of the present invention.

Claims

1. A semiconductor manufacturing intelligent predictive closed-loop control method, characterized in that: The method comprises the following steps: Acquire real-time FDC data of the machine, input the FDC data into a preset virtual measurement model, and output a first inline measurement parameter of the work-in-progress, where the first inline measurement parameter is a predicted value of an unmeasured inline measurement parameter; Inputting a third inline measurement parameter into a preset inline-Wat prediction model to output a predicted Watt value of the work-in-progress, wherein 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 Watt value exceeds the abnormal control range, the inline measurement parameters of the current process and the subsequent process of the product are corrected so that the predicted Watt value obtained by the inline-Wat prediction model after the correction of the inline measurement parameters does not exceed the abnormal control range; The step of obtaining real-time FDC data of the machine, inputting the FDC data into a preset virtual measurement model, and outputting the first inline measurement parameters of the work-in-progress includes: Acquiring historical data of the machine, the historical data including FDC data of the production equipment and a second inline measurement parameter, and preprocessing the historical data, the second inline measurement parameter being a measured inline measurement parameter; Constructing a virtual measurement model, and training and optimizing the virtual measurement model based on preprocessed historical data; Acquire the real-time FDC data of the machine, input the FDC data into the optimized virtual measurement model, and output the first inline measurement parameters of the work-in-progress; The step of pre-processing the historical data includes: determining whether each data point in the historical data exceeds a preset target range, and if so, removing the data point; if not, retaining the data point.

2. The semiconductor manufacturing intelligent prediction closed-loop control method according to claim 1, characterized in that: After the step of pre-processing 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 predicted residual value R, residual mean M, and residual standard deviation S of the virtual measurement model, where the predicted residual value R is the difference between the true inline measurement value and the first inline measurement parameter; Determine whether the residual value R is in the interval [M-3S, M+3S]. If not, remove the data point; if so, retain the data point.

3. The semiconductor manufacturing intelligent prediction closed-loop control method according to claim 1, characterized in that: The steps of constructing a virtual measurement model and training and optimizing the virtual measurement model according to the pre-processed historical data include: Constructing a data set from the preprocessed historical data, wherein the data set includes a training set and a test set; Establish a virtual measurement model, generate key hyperparameter combinations of the corresponding model on the training set, and cross-validate the key hyperparameter combinations one by one to obtain the mean absolute percentage error 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; Training the virtual measurement model on a training set using the optimal hyperparameters; The mean absolute percentage error score of the virtual measurement model on the test set is calculated, and the virtual measurement model with the smallest mean absolute percentage error score on the test set is selected as the optimal model.

4. The semiconductor manufacturing intelligent prediction closed-loop control method according to claim 3, characterized in that: The step of cross-validating the key hyperparameter combinations one by one includes: The training set is divided into K subsets, K-1 subsets are selected to train the virtual measurement model, and the remaining 1 subset is used to verify the virtual measurement model. The above process is repeated to obtain the virtual measurement model in the training set. The mean absolute percentage error score on .

5. The semiconductor manufacturing intelligent predictive 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 to output a predicted Watt value of the work-in-progress includes: Filter the third Inline measurement parameter; Calculate the correlation coefficient between the filtered third inline measurement parameter and the Watt value; Determine the contribution weights corresponding to the third inline measurement parameters at different levels, and calculate the comprehensive contribution weights of the third inline measurement parameters to the Watt value based on the correlation coefficients and the contribution weights; Based on the comprehensive contribution weight, the optimal Inline measurement parameter set is determined by a recursive feature method. Multiple prediction models are constructed according to the optimal Inline measurement parameter set. The mean absolute percentage error of the multiple prediction models on the test set is calculated. Difference score, the model with the smallest mean absolute percentage error score is selected as the Inline-Wat prediction model; The filtered third inline measurement parameter is input into the inline-Wat prediction model, and the predicted Watt value of the work-in-progress is output.

6. The semiconductor manufacturing intelligent prediction closed-loop control method according to claim 5, characterized in that: The step of determining the optimal Inline measurement parameter set based on the comprehensive contribution weight through a recursive feature method specifically includes: Determining an initial parameter set from a third Inline measurement parameter based on the comprehensive contribution weight; The initial parameter set is divided into a training set and a test set, and an xgboost model is constructed on the training set to train the initial parameter set; The mean absolute percentage error scores of all parameters in the initial parameter set in the test set are calculated, and the parameter set with the smallest mean absolute percentage error score is selected as the optimal inline measurement parameter set.

7. The semiconductor manufacturing intelligent predictive closed-loop control method according to claim 1, characterized in that: The third inline measurement parameters include: thickness parameters, size parameters, material parameters and product parameters.

8. A semiconductor manufacturing intelligent prediction closed-loop control system that implements the semiconductor manufacturing intelligent prediction closed-loop control method according to any one of claims 1 to 7, characterized in that: The system comprises: Data acquisition unit, used to obtain real-time FDC data of the machine; a first prediction unit, configured to input the FDC data into a preset virtual measurement model and output a first inline measurement parameter of the work-in-progress, where the first inline measurement parameter is a predicted value of an unmeasured inline measurement parameter; The second prediction unit is used to input the third inline measurement parameter into the preset inline-Wat prediction model and output the predicted Watt value of the work-in-progress. 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; The correction feedback unit is used to correct the inline measurement parameters of the current process and the subsequent process of the product when it is determined that the predicted Watt value exceeds the abnormal control range, so that the predicted Watt value obtained by the output of the inline-Wat prediction model after the correction of the inline measurement parameters does not exceed the abnormal control range.

9. A readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to implement the semiconductor manufacturing intelligent predictive closed-loop control method according to any one of claims 1 to 7.

Images