Wafer-level test yield prediction and process optimization method and system based on big data

Abstract

The invention provides a wafer-level test yield prediction and process optimization method and system based on big data, and relates to the technical field of wafer-level tests.The wafer-level test yield prediction and process optimization method comprises the steps that wafer manufacturing process data is obtained, a time-sequence-sensitive multi-modal data fusion network is constructed, a LSTM and GNN combined mixed architecture is adopted, and a multi-modal data fusion network is constructed; self-adaptive weight distribution is carried out through an attention mechanism to predict the yield; and when the predicted yield is lower than a threshold value, triggering a layered progressive process optimization system, and adjusting process parameters through an equipment parameter optimization layer and a process level optimization layer. According to the invention, the accurate prediction of the wafer test yield and the automatic optimization of the process parameters are realized, the production efficiency is improved, and the manufacturing cost is reduced.

Description

Technical Field

[0001] The present invention relates to wafer-level testing technology, and particularly to a method and system for predicting wafer-level testing yield and optimizing processes based on big data. Background Art

[0002] With the continuous development of integrated circuit manufacturing processes, the wafer manufacturing process has become increasingly complex. The process nodes continue to shrink, and the requirements for manufacturing precision and yield control have also increased accordingly. Wafer testing yield is a key indicator for measuring the quality of chip manufacturing, directly affecting product cost and market competitiveness. Traditional wafer testing mainly relies on electrical performance testing in the later stage of manufacturing, and it is unable to detect and predict potential yield problems in a timely manner.

[0003] In recent years, with the application of big data and artificial intelligence technologies in the semiconductor manufacturing field, data-driven yield prediction and process optimization methods have gradually received attention. The massive data generated during the wafer manufacturing process includes various modal information such as process parameter data, equipment status data, and defect detection data. These data contain rich yield-related features. Effectively using these data for yield prediction and process optimization can significantly improve manufacturing efficiency and product quality.

[0004] Current research mainly focuses on using machine learning methods to analyze wafer testing data, establish a yield prediction model, and adjust process parameters based on the prediction results. However, how to effectively integrate multi-modal data, capture complex temporal and spatial features, and how to systematically perform multi-level process optimization are still important challenges in this field.

[0005] Existing methods for predicting wafer-level testing yield and optimizing processes mainly have the problem of limited data fusion ability. Most methods only focus on a single type of data source, such as process parameters or defect images, and it is difficult to effectively integrate multi-modal data such as process parameters, equipment status, and defect detection. Especially when dealing with the temporal characteristics and spatial topological relationships of these data, there is a lack of an effective fusion mechanism, resulting in the prediction model being unable to fully utilize multi-dimensional information and affecting prediction accuracy.

[0006] Secondly, there are limitations in the construction of existing yield prediction models. Traditional machine learning methods such as support vector machines or simple neural networks are difficult to capture both the temporal dependence and spatial correlation in the wafer manufacturing process simultaneously. Although some research attempts to use deep learning methods, most use a single network structure, such as only using CNN or LSTM, and are unable to model temporal changes and spatial features simultaneously, resulting in insufficient ability to predict yield fluctuations in complex process environments.

[0007] In addition, existing process optimization methods often adopt a single-level optimization strategy and lack a systematic and hierarchical optimization framework. Most methods either focus on local adjustment of individual equipment parameters or conduct coarse-grained process route optimization, making it difficult to achieve integrated optimization from equipment parameters to the overall process flow. At the same time, existing optimization algorithms usually adopt a single optimization method, such as gradient descent or genetic algorithm, lacking the co-evolution ability of multiple optimization algorithms and making it difficult to find the global optimal solution in a complex parameter space, which limits the effect of process optimization. Summary of the Invention

[0008] The embodiments of the present invention provide a method and system for predicting wafer-level test yield and optimizing process based on big data, which can solve the problems in the prior art.

[0009] In the first aspect of the embodiments of the present invention,

[0010] A method for predicting wafer-level test yield and optimizing process based on big data is provided, including:

[0011] Construct a time-series sensitive multi-modal data fusion network. The multi-modal data fusion network adopts a hybrid architecture combining long short-term memory network (LSTM) and graph neural network (GNN). The LSTM is used to capture the time-series characteristics of process parameter data, equipment status data, and defect detection data. The GNN is used to establish the spatial topological relationship of test yield data. An attention mechanism is used to adaptively allocate weights to the spatial features corresponding to the time-series characteristics and spatial topological relationship, so as to predict the wafer test yield and obtain a predicted yield value.

[0012] When the predicted yield value is lower than a preset yield threshold, trigger the process optimization process.

[0013] Construct a hierarchical progressive process optimization system. The hierarchical progressive process optimization system includes an equipment parameter optimization layer and a process-level optimization layer. The equipment parameter optimization layer uses the co-evolution method of genetic algorithm and particle swarm optimization algorithm to locally optimize a single process parameter and group optimize the parameter combinations of multiple related devices. The overall manufacturing process is globally optimized at the process-level optimization layer, and the optimization results of each optimization layer are comprehensively evaluated and dynamically adjusted based on the Bayesian optimization framework.

[0014] Adjust the process parameters according to the optimization plan generated by the hierarchical progressive process optimization system to generate a process optimization instruction.

[0015] The process parameter data includes etching depth, lithography alignment accuracy, metal deposition thickness, and doping concentration; the equipment status data includes chamber pressure, RF power, gas flow rate, and temperature; the defect detection data includes defect size, defect type, defect density, and defect distribution; the test yield data includes functional test yield, parameter test yield, and reliability test yield.

[0016] Through the attention mechanism, an adaptive weight assignment is performed on the spatial features corresponding to the temporal features and the spatial topological relationship to predict the wafer test yield, and the predicted yield value is obtained, including:

[0017] The long short-term memory network is used to extract features from the process parameter data, the equipment status data, and the defect detection data, and selective memory is performed through the forget gate, input gate, and output gate to obtain temporal features;

[0018] Based on the test yield data, a graph structure is constructed. The bad points on the wafer are used as the nodes of the graph structure, and the spatial correlation relationship between adjacent bad points is used as the edges of the graph structure. The graph neural network is used to extract features from the graph structure, and spatial features are obtained through node feature aggregation;

[0019] Taking the temporal features as the query vector, taking the spatial features as the key vector and value vector, calculating the similarity between the query vector and the key vector to obtain the attention weight, and performing weighted summation on the value vector based on the attention weight to obtain the fused feature;

[0020] The fused feature is input into the prediction layer, and the predicted yield value is obtained through non-linear transformation. The prediction layer uses the sigmoid activation function to map the prediction output to between 0 and 1.

[0021] Through the co-evolution method of the genetic algorithm and the particle swarm optimization algorithm, local optimization of a single process parameter is performed in the parameter-level optimization layer, including:

[0022] Encoding the process parameter to be optimized as a real chromosome to generate a parameter encoding vector. The parameter encoding vector contains n process parameter components, and corresponding parameter upper limit values and parameter lower limit values are set for each process parameter component; constructing a fitness function with the yield improvement amount, process cost, and parameter fluctuation degree as the optimization objectives, and performing weighted combination on the yield improvement amount, the process cost, and the parameter fluctuation degree through weight coefficients;

[0023] Generating an initial genetic algorithm population and an initial particle swarm population. The initial genetic algorithm population contains N chromosome individuals, and the initial particle swarm population contains M particle individuals;

[0024] Perform a selection operation on the genetic algorithm population. Use the roulette wheel selection method to calculate the selection probability based on the individual fitness value, and select high-quality individuals to enter the mating pool; perform an arithmetic crossover operation on the individuals in the mating pool, generate offspring individuals through the crossover coefficient, and perform a Gaussian mutation operation on the offspring individuals;

[0025] Update the velocity and position information of each particle in the particle swarm, and adjust the particle position according to the updated velocity;

[0026] Exchange the optimal individual information between the genetic algorithm population and the particle swarm population every preset number of generations. Inject the optimal individual information of the genetic algorithm population into the particle swarm population, and inject the optimal individual information of the particle swarm population into the genetic algorithm population;

[0027] Use a dynamic adjustment coefficient to perform weighted combination on the fitness of the genetic algorithm population and the particle swarm population to obtain a mixed fitness value; perform local fine search in the neighborhood of the currently obtained optimal solution, and fine-tune the optimal solution through the search step size and the search direction vector; judge the convergence of the mixed fitness value. When the difference between the mixed fitness values of adjacent generations is less than the preset fitness threshold, output the optimal process parameter combination.

[0028] Perform global optimization on the overall manufacturing process at the process-level optimization layer, and comprehensively evaluate and dynamically adjust the optimization results of each optimization layer based on the Bayesian optimization framework, including:

[0029] Construct a directed acyclic graph of the process flow. The directed acyclic graph of the process flow includes a set of process nodes, a set of process dependency edges, and a node weight matrix. Each process node in the set of process nodes contains a set of process parameters, processing time, quality indicators, and cost factors;

[0030] Construct a multi-objective comprehensive evaluation function based on the directed acyclic graph of the process flow. The multi-objective comprehensive evaluation function includes a manufacturing cycle term, a product yield term, and a resource utilization rate term. Calculate the manufacturing cycle term according to the processing time in the process node, calculate the product yield term according to the quality indicator, calculate the resource utilization rate term according to the cost factor, and perform weighted combination through a dynamic weight coefficient;

[0031] Establish a Gaussian process regression model based on the multi-objective comprehensive evaluation function, and construct a Bayesian optimization framework using the Matérn kernel function; optimize each optimization layer based on the Bayesian optimization framework to obtain the optimization results of each optimization level. The optimization results include an optimal solution vector, an objective function value, and a confidence level.

[0032] The method further includes:

[0033] Construct a combined evaluation index to evaluate the optimization result, establish an adaptive exploration mechanism according to the evaluation result, dynamically select sampling points through an exploration-exploitation balance factor, and the exploration-exploitation balance factor is adaptively adjusted as the optimization process progresses;

[0034] Monitor the process parameters in the wafer manufacturing process in real time. When the drift amount of the process parameters relative to the preset reference value exceeds the preset drift threshold, generate a compensation vector based on a proportional-integral-derivative controller;

[0035] Apply the compensation vector to the current process parameters to obtain a compensated process parameter combination, and input the compensated process parameter combination into the multi-objective comprehensive evaluation function;

[0036] When the output value of the multi-objective comprehensive evaluation function meets the convergence condition and the spatial distribution of the compensated process parameter combination meets the aggregation condition, determine the final optimization plan.

[0037] In the second aspect of the embodiments of the present invention,

[0038] Provide a wafer-level test yield prediction and process optimization system based on big data, including:

[0039] The first unit is used to obtain process parameter data, equipment status data, defect detection data, and test yield data in the wafer manufacturing process;

[0040] The second unit is used to construct a time-series sensitive multi-modal data fusion network. The multi-modal data fusion network adopts a hybrid architecture combining a long short-term memory network (LSTM) and a graph neural network (GNN). Capture the time-series characteristics of process parameter data, equipment status data, and defect detection data through LSTM, establish the spatial topological relationship of test yield data using GNN, and perform adaptive weight allocation on the spatial characteristics corresponding to the time-series characteristics and the spatial topological relationship through an attention mechanism to predict the wafer test yield and obtain the predicted yield value;

[0041] The third unit is used to trigger the process optimization process when the predicted yield value is lower than the preset yield threshold;

[0042] The fourth unit is used to construct a hierarchical progressive process optimization system. The hierarchical progressive process optimization system includes an equipment parameter optimization layer and a process-level optimization layer. The equipment parameter optimization layer uses a co-evolution method of a genetic algorithm and a particle swarm optimization algorithm to perform local optimization on a single process parameter and group optimization on the parameter combinations of multiple related devices. Perform global optimization on the overall manufacturing process in the process-level optimization layer, and perform comprehensive evaluation and dynamic adjustment on the optimization results of each optimization layer based on the Bayesian optimization framework;

[0043] The fifth unit is configured to adjust process parameters according to the optimization scheme generated by the hierarchical progressive process optimization system, and generate a process optimization instruction.

[0044] In the third aspect of the embodiments of the present invention,

[0045] There is provided an electronic device, including:

[0046] A processor;

[0047] A memory for storing instructions executable by the processor;

[0048] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0049] In the fourth aspect of the embodiments of the present invention,

[0050] There is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0051] The beneficial effects of this application are as follows:

[0052] By constructing a time-sensitive multi-modal data fusion network and combining the hybrid architecture of LSTM and GNN, it is possible to simultaneously capture the time series characteristics of process parameters, equipment status, and defect detection data, as well as the spatial topological relationship of test yield data, and perform adaptive weight allocation through the attention mechanism, achieving high-precision prediction of wafer test yield and improving the accuracy and reliability of the prediction model.

[0053] A hierarchical progressive process optimization system is adopted, which includes an equipment parameter optimization layer and a process-level optimization layer. Multi-level optimization is achieved through the co-evolution method of genetic algorithm and particle swarm optimization algorithm. It can not only perform local optimization on a single process parameter, but also perform population optimization on the parameter combinations of multiple related devices, and at the same time achieve global optimization of the overall manufacturing process at the process level, greatly improving the efficiency and effect of process optimization.

[0054] Based on the Bayesian optimization framework, comprehensive evaluation and dynamic adjustment of the optimization results of each optimization layer are carried out, realizing the adaptability and robustness of the optimization process. It can quickly respond according to the actual production situation and generate the best process optimization scheme, reducing the wafer manufacturing cost, improving the product yield and production efficiency, and providing a technically practical solution for the semiconductor manufacturing industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a schematic flowchart of the method for wafer-level test yield prediction and process optimization based on big data in the embodiments of the present invention;

[0056] Figure 2 Schematic diagram for comparing the yield detection capabilities under different defect types in the embodiments of the present invention;

[0057] Figure 3 Schematic diagram for visualizing the weight distribution of the attention mechanism in the technical solution of the embodiments of the present invention;

[0058] Figure 4 Experimental data comparison data table for the process parameter optimization method in the embodiments of the present invention;

[0059] Figure 5 Manufacturing process global optimization performance comparison data table for the embodiments of the present invention. Detailed implementation manners

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part rather than all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0061] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0062] Figure 1 Schematic flowchart of the method for predicting the wafer-level test yield and optimizing the process based on big data in the embodiments of the present invention, as Figure 1 shown, the method includes:

[0063] Obtain process parameter data, equipment status data, defect detection data, and test yield data during the wafer manufacturing process;

[0064] Construct a time-series sensitive multi-modal data fusion network. The multi-modal data fusion network adopts a hybrid architecture combining a long short-term memory network (LSTM) and a graph neural network (GNN). Capture the time-series features of process parameter data, equipment status data, and defect detection data through LSTM, establish the spatial topological relationship of test yield data using GNN, and perform adaptive weight allocation on the spatial features corresponding to the time-series features and the spatial topological relationship through an attention mechanism to predict the wafer test yield and obtain a predicted yield value;

[0065] When the predicted yield value is lower than a preset yield threshold, trigger the process optimization process;

[0066] Construct a hierarchical progressive process optimization system. The hierarchical progressive process optimization system includes an equipment parameter optimization layer and a process-level optimization layer. The equipment parameter optimization layer performs local optimization on a single process parameter and population optimization on the parameter combinations of multiple related devices through the co-evolution method of the genetic algorithm and the particle swarm optimization algorithm. The process-level optimization layer performs global optimization on the overall manufacturing process and comprehensively evaluates and dynamically adjusts the optimization results of each optimization layer based on the Bayesian optimization framework;

[0067] Adjust the process parameters according to the optimization scheme generated by the hierarchical progressive process optimization system to generate a process optimization instruction.

[0068] In an optional implementation manner, the process parameter data includes etching depth, lithography alignment accuracy, metal deposition thickness, and doping concentration; the equipment status data includes chamber pressure, radio frequency power, gas flow rate, and temperature; the defect detection data includes defect size, defect type, defect density, and defect distribution; the test yield data includes functional test yield, parameter test yield, and reliability test yield.

[0069] The system is provided with a special data acquisition module for acquiring four types of key data. The process parameter data acquisition sub-module adopts a distributed sensor network, and multiple high-precision sensors are set at key nodes of wafer processing. For example, in the etching process, a laser interferometer is used to monitor the etching depth in real time, with an accuracy of up to ±0.5 nm; in the lithography alignment process, a CCD image recognition system is used to monitor the alignment accuracy, which is controlled within ±5 nm; during the metal deposition process, a quartz crystal microbalance is used to monitor the film thickness, with an accuracy of ±0.1 nm; the doping concentration is measured by a secondary ion mass spectrometer, and the detectable range is 1E15 to 1E21 / cm³.

[0070] The equipment status data acquisition sub-module mainly monitors the operating status of production equipment. The chamber pressure is monitored in real time by a capacitive pressure sensor, and within the working pressure range of 0.5 - 500 mTorr, the accuracy can reach ±0.1 mTorr; the radio frequency power is monitored by a directional coupler, and for a device with a nominal power of 1000 W, the control accuracy is within ±5 W; the gas flow rate is monitored by a mass flow controller, and within the range of 10 - 1000 sccm, the control accuracy is ±1% of the set value; the temperature is monitored by an infrared sensor array, and the temperature range is from room temperature to 1200 °C, with an accuracy of ±0.5 °C.

[0071] The defect detection data acquisition sub-module combines a high-resolution scanning electron microscope and an optical defect inspection system. The defect size measurement range is from 10nm to 10μm, with an accuracy of ±1nm; the defect types are automatically classified into 10 types such as particle contamination, scratches, bridging, residues, etc. through image recognition algorithms; the defect density calculates the number of defects per unit area, usually 0.01 - 10 per cm²; the defect distribution shows the distribution of defects on the wafer through a heat map, which can be subdivided into 1mm×1mm blocks.

[0072] The test yield data acquisition sub-module includes three subsystems. The functional test yield is collected by an automatic test equipment (ATE), and the test coverage rate reaches over 98%; the parametric test yield collects key electrical parameters such as threshold voltage, leakage current, breakdown voltage, etc. through a parameter analyzer, and the measurement accuracy can reach ±0.5% of the set value; the reliability test yield is obtained through methods such as aging test, temperature cycle test, and high temperature and high humidity test. The aging test condition is 125℃ for 168 hours, and the temperature cycle test condition is -40℃ to 125℃ for 500 cycles.

[0073] The data preprocessing module first cleans and standardizes the collected raw data. For possible outliers, the system uses the moving average method for smoothing, with a moving window size of 5 data points; for missing values, the nearest neighbor interpolation method is used to fill them, with a maximum allowable interpolation distance of 3 data points; data standardization uses the Z-Score method to enable different types of data to be compared under the same standard.

[0074] The correlation analysis module realizes the mining of the relationships between different parameters. For example, regarding the relationship between etching depth and RF power, the system finds that within the power range of 300 - 500W, the etching depth is positively correlated with the power, with a correlation coefficient of 0.85; when the chamber pressure changes within the range of 5 - 10mTorr, the etching uniformity decreases by 17%; the metal deposition thickness and temperature show a linear relationship within the range of 250 - 300℃, and for every 5℃ increase, the thickness increases by approximately 2.5nm.

[0075] The anomaly detection module uses statistical analysis methods to identify process anomalies. When a certain parameter deviates from the set value by more than 3 standard deviations, the system triggers an alarm; for the situation where 5 consecutive measurement points are within the control limits but show a monotonic trend, the system issues a warning message; when the defect density suddenly rises from 0.05 per cm² to 0.15 per cm², the system automatically analyzes the possible reasons and provides intervention suggestions.

[0076] The yield prediction module constructs a prediction model based on historical data. This module can predict the final yield at the initial stage of production based on the current process parameters and equipment status. In practical applications, when the etching depth is controlled within 200 ± 5 nm, the lithography alignment accuracy is within ± 3 nm, the metal thickness is 300 ± 3 nm, and the doping concentration is 1E18 ± 5% / cm³, the functional test yield can reach over 95%; when the defect density is controlled below 0.03 per cm², the parametric test yield can be increased to 97%.

[0077] The parameter optimization module determines the optimal process window by analyzing historical production data. For example, for a specific 90nm logic process, the optimal etching parameter window is: RF power 450 ± 10 W, chamber pressure 7.5 ± 0.3 mTorr, gas flow ratio CF4:O2 = 50:10 ± 2 sccm, temperature 65 ± 1 °C; for the metal deposition process, the optimal parameter window is: deposition temperature 275 ± 3 °C, chamber pressure 3 ± 0.1 mTorr, power 200 ± 5 W, argon flow 50 ± 1 sccm.

[0078] The real-time regulation module makes timely adjustments according to the monitored parameter offsets. When it is detected that the etching depth is too shallow, the system automatically extends the etching time from 60 seconds to 63 seconds; when the monitored metal thickness is too thin, the deposition rate is adjusted from 3 nm / minute to 3.2 nm / minute; when the doping concentration is too low, the ion implantation energy is increased from 80 keV to 82 keV. These adjustments are all automatically completed through the equipment interface without manual intervention.

[0079] The defect control module takes corresponding corrective measures for the detected defect types. For defects of particle contamination type, the system automatically increases the number of wafer cleaning times from the standard 2 times to 3 times, and increases the ultrasonic power of the last cleaning from 300 W to 350 W; for residual defects caused by photoresist residues, the system extends the development time from 60 seconds to 65 seconds and increases the developer temperature from 21 °C to 22 °C.

[0080] The yield improvement feedback module correlates and analyzes the yield changes with process parameter adjustments. In a certain production activity, by improving the etching uniformity from ± 3% to ± 1.5%, the functional test yield is increased from 92% to 95.5%; by optimizing the lithography alignment accuracy from ± 5 nm to ± 3 nm, the parametric test yield is increased from 94% to 96.3%; by reducing the impurity content during metal deposition from 100 ppm to 20 ppm, the reliability test yield is increased from 90% to 94.8%.

[0081] Through the implementation of the above system, it is possible to accurately monitor the entire integrated circuit manufacturing process, quickly identify and solve problems in the production process, and ultimately achieve continuous improvement in product yield, significant improvement in production efficiency, and effective guarantee of product consistency.

[0082] Figure 2 This is a schematic diagram comparing the yield detection capabilities under different defect types according to an embodiment of the present invention:

[0083] This figure compares the performance of three different solutions (this technology solution, the traditional model solution, and the lightweight verification solution) on seven different metrics. The three solutions are represented by solid triangles, dashed dots, and dotted squares, respectively.

[0084] The data shows that the proposed technology solution (solid triangles) generally performs best across all metrics, particularly for point defects (0.5μm), achieving approximately 97 points. Line defects (1.2μm) and area defects (5μm) both approach 95 points. It also maintains high scores between 92 and 95 for metrics such as lithography offset (0.2μm), metal shorts, underetching (20%), and particulate contamination (0.3μm). The traditional model solution (dashed circles) performs second best overall, with scores ranging from 86 to 92 across all metrics, generally 3-5 points lower than the proposed technology solution. The light verification solution (dotted squares) performs the worst, with scores ranging from 82 to 88 across all metrics, approximately 8-10 points lower than the proposed technology solution.

[0085] The figure clearly shows the advantages and disadvantages of the three solutions in terms of various technical indicators. This technical solution shows obvious technical advantages in all seven evaluation dimensions, especially in defect detection, which fully demonstrates the advanced nature and practical value of the solution.

[0086] In semiconductor manufacturing, traditional defect detection methods rely primarily on fixed detection models and preset parameter thresholds. These methods often suffer from insufficient detection accuracy, high rates of missed detections, and false positives when faced with complex and changing manufacturing environments and different types of defects. While lightweight verification schemes offer the advantage of fast processing speed, they simplify detection algorithms and feature extraction, resulting in lower reliability and accuracy.

[0087] By introducing adaptive feature extraction and multi-scale analysis techniques, combined with deep learning algorithms to model the characteristics of different defect types, this solution enables accurate identification of all types of defects in the semiconductor manufacturing process. This solution specifically targets defect types of varying scales, such as point defects, line defects, and area defects, and employs a hierarchical detection strategy, ensuring comprehensive detection while improving identification accuracy.

[0088] This solution innovatively adopts a dynamic threshold adjustment mechanism, automatically optimizing the detection parameters according to actual process parameters and environmental conditions, effectively solving the problem that traditional fixed-threshold solutions are difficult to adapt to process fluctuations. At the same time, by introducing multi-dimensional feature fusion technology, the detection ability for various defect types such as lithography offset, metal short circuit, etching deficiency, and particle contamination is improved.

[0089] This solution has achieved significant improvements in aspects such as the accuracy, reliability, and adaptability of defect detection, being significantly superior to traditional detection solutions and lightweight verification solutions. Especially in defect detection under complex process environments, it performs more prominently, greatly reducing the missed detection rate and false detection rate, improving the detection efficiency, and providing a more reliable technical guarantee for semiconductor manufacturing quality control.

[0090] This solution also has good scalability and compatibility, and can be flexibly adjusted and optimized according to different manufacturing process requirements, providing a more comprehensive and efficient solution for quality monitoring in the semiconductor manufacturing process. These improvements not only increase production efficiency but also provide strong support for improving product yield.

[0091] In an alternative embodiment, an adaptive weight allocation is performed on the spatial features corresponding to the temporal features and the spatial topological relationship through an attention mechanism to predict the wafer test yield, and the obtained predicted yield value includes:

[0092] Use a long short-term memory network to extract features from the process parameter data, the equipment status data, and the defect detection data, and perform selective memory through the forget gate, input gate, and output gate to obtain temporal features;

[0093] Based on the test yield data, construct a graph structure, take the bad points on the wafer as the nodes of the graph structure, take the spatial correlation relationship between adjacent bad points as the edges of the graph structure, and use a graph neural network to extract features from the graph structure, and obtain spatial features through node feature aggregation;

[0094] Take the temporal features as the query vector, take the spatial features as the key vector and value vector, calculate the similarity between the query vector and the key vector to obtain the attention weight, and perform weighted summation on the value vector based on the attention weight to obtain the fused feature;

[0095] Input the fused feature into the prediction layer, and obtain the predicted yield value through a non-linear transformation. The prediction layer uses a sigmoid activation function to map the prediction output to between 0 and 1.

[0096] Collect and preprocess various types of data in the wafer manufacturing process, including process parameter data, equipment status data, defect detection data, and test yield data. Process parameter data includes parameters such as temperature, pressure, and gas flow; equipment status data includes information such as equipment running time and maintenance records; defect detection data includes information such as the location, size, and type of defects on the wafer surface; and test yield data records the final test results of each wafer.

[0097] Preprocess the collected data, including operations such as outlier handling, missing value filling, and normalization. For example, for the temperature parameter, the raw data in the range of 250 - 350 degrees Celsius can be normalized to the range of 0 - 1; for missing equipment status records, they can be filled with the average value of the previous and next time points.

[0098] Use the Long Short-Term Memory network (LSTM) to extract temporal features from process parameter data, equipment status data, and defect detection data. The LSTM network consists of an input layer, an LSTM layer, and an output layer. The LSTM layer contains multiple LSTM units, and each unit contains a forget gate, an input gate, and an output gate.

[0099] The forget gate determines the historical information to be discarded. For example, when it is detected that the equipment has been maintained, some of the previously accumulated equipment status information may no longer be relevant. At this time, the forget gate will output a value close to 0, indicating to discard this information. Specifically, the forget gate receives the input at the current moment and the hidden state at the previous moment, and outputs a value between 0 and 1 through the sigmoid function, which is used to control the retention ratio of the unit state at the previous moment.

[0100] The input gate controls the update degree of the new information at the current moment. For example, when the newly introduced photoresist changes, the input gate will output a value close to 1, indicating that this information is important and needs to be remembered. The input gate receives the input at the current moment and the hidden state at the previous moment, outputs a value between 0 and 1 through the sigmoid function, and generates a new candidate unit state through the tanh function.

[0101] The output gate controls the degree to which the current unit state affects the output. For example, when a critical process parameter anomaly is detected, the output gate will output a higher value to reflect this situation. The output gate receives the input at the current moment and the hidden state at the previous moment, outputs a value between 0 and 1 through the sigmoid function, and multiplies it with the unit state processed by the tanh function to obtain the final output.

[0102] After being processed by the LSTM network, a temporal feature vector with a dimension of 128 is obtained, which captures the patterns of process parameters, equipment status, and defect detection data changing over time.

[0103] Construct a graph structure based on the test yield data. Consider the defective points on the wafer as the nodes of the graph structure, and each node contains the location information and defect type of the defective point. Take the spatial correlation relationship between adjacent defective points as the edges of the graph structure, and define that there is an edge connection when the distance between two defective points is less than a threshold (e.g., 100 microns).

[0104] Use a graph neural network (GNN) to extract spatial features from the constructed graph structure. GNN realizes the aggregation and update of node features through a message passing mechanism. For each node, first collect the feature information of its neighbor nodes, and then update its own features based on the collected information.

[0105] The initial feature of each node is encoded by its position coordinates and defect type. For example, for a particle defect located at the wafer coordinates (10.5mm, 15.2mm), it can be encoded as an initial feature vector. The aggregation of neighbor node features is achieved through weighted summation, and the weights can be determined according to the distance between the two nodes. The closer the distance, the greater the weight. For example, when the distance between two defective points is 50 microns, the corresponding weight may be 0.8; when the distance is 90 microns, the weight may drop to 0.4.

[0106] After 3 rounds of message passing, the features of each node already contain the information of its neighborhood structure. Through a global pooling operation, the features of all nodes are aggregated into a unified graph representation, obtaining a spatial feature vector with a dimension of 96, which captures the spatial distribution pattern of the defective points on the wafer.

[0107] Implement an attention mechanism to fuse the temporal features and spatial features. Take the temporal features as the query vector, and the spatial features as the key vector and value vector. First, map the query vector, key vector, and value vector to the same attention space through a linear transformation.

[0108] Obtain the attention weights by calculating the similarity between the query vector and the key vector. For example, when the temporal features show an abnormal increase in temperature at a certain moment, and the spatial features show concentrated defects in the edge area of the wafer, there may be a high correlation between these two features, and the corresponding attention weight will be relatively high, such as 0.75. The attention weights are normalized after softmax processing to ensure that the sum of all weights is 1.

[0109] Perform weighted summation on the value vector based on the obtained attention weights to obtain the fused features. This fused feature contains both the temporal information of process parameters, equipment status, and defect detection, as well as the spatial topological information of the defective point distribution, with a dimension of 128.

[0110] The fused features are input into the prediction layer for yield prediction. The prediction layer consists of two fully connected layers, with a ReLU activation function added in the middle to increase the non-linear expression ability. The first layer maps the 128-dimensional fused features to 64-dimensional hidden features, and the second layer maps the hidden features to a 1-dimensional prediction output.

[0111] The sigmoid activation function is used to map the prediction output to a value between 0 and 1, representing the predicted yield value. For example, if the prediction output is 0.873, it means the predicted yield of the wafer is 87.3%. In practical applications, when the yield prediction value is lower than a specific threshold (such as 0.8), the warning mechanism can be triggered to prompt the engineer to check for possible process anomalies.

[0112] Through verification on actual production data, the prediction accuracy of this method reaches 92.5%, which is 7.3% and 5.8% higher than the traditional methods that only use temporal features or only use spatial features respectively. It can effectively guide the optimization of the wafer manufacturing process and improve production efficiency and product yield.

[0113] Figure 3 This is the visualization schematic diagram of the attention mechanism weight assignment in the embodiment of the present invention and the technical solution of the present application:

[0114] This figure shows the experimental analysis results of an attention mechanism, including the attention weight assignment matrix in two dimensions of temporal features and spatial features, as well as the comparison data of three key performance indicators.

[0115] In the attention weight assignment matrix of temporal features, the values generally show a decreasing trend from the upper left corner to the lower right corner, with the highest weight reaching 0.98 and the lowest weight being 0.41. The weight distribution shows that the attention focus in the time series mainly concentrates on the first half of the sequence, indicating that the model pays more attention to recent feature information when processing temporal information.

[0116] The attention weight assignment matrix of spatial features shows an increasing trend from the lower left corner to the upper right corner, with the highest weight reaching 0.96 and the lowest weight being 0.41. This distribution pattern indicates that the model has different degrees of attention to features at different positions in the spatial dimension and is biased towards the spatial features in the upper right region.

[0117] In terms of performance indicators, the technical solution of the present application shows significant advantages: the attention focusing efficiency reaches 89.7%, which is significantly improved compared with 63.2% of the non-attention solution; the feature fusion accuracy reaches 95.4%, which is much higher than 82.1% of the non-attention solution; in terms of prediction performance, it is improved by 7.3% compared with LSTM and 7.0% compared with GNN. These data fully illustrate the effectiveness of this attention mechanism in improving the model performance, especially in feature extraction and information fusion.

[0118] These results comprehensively demonstrate the rationality of the attention allocation strategy of this technical solution in both the temporal and spatial dimensions, as well as the performance improvement brought about in practical applications, proving the superiority of this solution in dealing with complex spatio-temporal features.

[0119] In an alternative embodiment, through the co-evolution method of the genetic algorithm and the particle swarm optimization algorithm, local optimization of a single process parameter at the parameter-level optimization layer includes:

[0120] Encoding the process parameter to be optimized as a real-number chromosome to generate a parameter encoding vector, where the parameter encoding vector contains n process parameter components, and corresponding upper and lower parameter limit values are set for each process parameter component; constructing a fitness function with the yield improvement amount, process cost, and parameter fluctuation degree as optimization objectives, and performing weighted combination on the yield improvement amount, the process cost, and the parameter fluctuation degree through weight coefficients;

[0121] Generating an initial genetic algorithm population and an initial particle swarm population, where the initial genetic algorithm population contains N chromosome individuals, and the initial particle swarm population contains M particle individuals;

[0122] Performing a selection operation on the genetic algorithm population, calculating the selection probability according to the individual fitness value using the roulette wheel selection method, and selecting high-quality individuals to enter the mating pool; performing an arithmetic crossover operation on the individuals in the mating pool, generating offspring individuals through the crossover coefficient, and performing a Gaussian mutation operation on the offspring individuals;

[0123] Updating the velocity and position information of each particle in the particle swarm, and adjusting the particle position according to the updated velocity;

[0124] Exchanging the optimal individual information between the genetic algorithm population and the particle swarm population every preset number of generations, injecting the optimal individual information of the genetic algorithm population into the particle swarm population, and injecting the optimal individual information of the particle swarm population into the genetic algorithm population;

[0125] Using a dynamic adjustment coefficient to perform weighted combination on the fitness of the genetic algorithm population and the particle swarm population to obtain a mixed fitness value; performing local fine search in the neighborhood of the currently obtained optimal solution, and fine-tuning the optimal solution through the search step size and the search direction vector; judging the convergence of the mixed fitness value, and when the difference between the mixed fitness values of two adjacent generations is less than the preset fitness threshold, outputting the optimal process parameter combination.

[0126] Encode the process parameters to be optimized as real - valued chromosomes to generate a parameter encoding vector. For example, for the etching process, key parameters such as power, pressure, gas flow rate, and chamber temperature can be selected for optimization. Suppose there are 5 process parameters to be optimized, then the parameter encoding vector is [P1, P2, P3, P4, P5], where P1 represents the power parameter, and its value range is [800W, 1200W]; P2 represents the chamber pressure, and its value range is [10mTorr, 50mTorr]; P3 represents the main gas flow rate, and its value range is [100sccm, 200sccm]; P4 represents the auxiliary gas flow rate, and its value range is [20sccm, 60sccm]; P5 represents the chamber temperature, and its value range is [20℃, 80℃].

[0127] Construct a fitness function with the yield improvement amount, process cost, and parameter fluctuation degree as the optimization objectives. The yield improvement amount is defined as the difference between the yield after optimization and the yield before optimization; the process cost includes energy consumption cost, material cost, and time cost; the parameter fluctuation degree represents the deviation degree of the optimized parameters from the original parameters. The fitness function combines these three indicators through weight coefficients. For example, the weight coefficients can be taken as 0.6, 0.3, and 0.1 respectively, with the yield improvement as the main objective, while considering cost and stability.

[0128] Generate an initial genetic algorithm population and an initial particle swarm population. The initial genetic algorithm population contains 40 chromosome individuals, and each individual is a vector containing 5 process parameters, and the parameter values are randomly generated within their respective value ranges. For example, the first individual may be [950W, 25mTorr, 150sccm, 35sccm, 50℃]. The initial particle swarm population contains 30 particle individuals, which are also randomly generated, and an initial velocity vector is specified for each particle. For example, the initial velocity of particle 1 may be [5W / generation, 2mTorr / generation, 3sccm / generation, 1sccm / generation, 2℃ / generation].

[0129] When operating on the genetic algorithm population, the roulette wheel selection method is first adopted. Calculate the fitness values of 40 individuals. Assuming the highest fitness is 0.85 and the lowest fitness is 0.32, then the probability that an individual with a fitness of 0.85 is selected is approximately 2.66 times that of an individual with a fitness of 0.32. Select 20 pairs of parent individuals to enter the mating pool, and then perform arithmetic crossover operations. For example, for two parent individuals [950W, 25mTorr, 150sccm, 35sccm, 50℃] and [1000W, 30mTorr, 160sccm, 40sccm, 60℃], when using a crossover coefficient of 0.7 for crossover, the first offspring individual is [965W, 26.5mTorr, 153sccm, 36.5sccm, 53℃]. Perform Gaussian mutation operations on the offspring individuals, set the mutation probability to 0.1, and the mutation amplitude to 5% of the parameter range. For example, when mutating the power parameter, it may change from 965W to 980W.

[0130] For the update of the particle swarm, a classical velocity-position update mechanism is adopted. Set the inertia weight to 0.7, the cognitive coefficient to 1.5, and the social coefficient to 1.5. For example, if the current position of a certain particle is [980W, 28mTorr, 155sccm, 38sccm, 55℃], the current velocity is [5W / generation, 2mTorr / generation, 3sccm / generation, 1sccm / generation, 2℃ / generation], the individual historical best position is [990W, 30mTorr, 158sccm, 40sccm, 58℃], and the global best position is [1000W, 32mTorr, 160sccm, 42sccm, 60℃], then the updated velocity may become [8.2W / generation, 3.5mTorr / generation, 4.1sccm / generation, 2.8sccm / generation, 3.7℃ / generation], and the new position is [988.2W, 31.5mTorr, 159.1sccm, 40.8sccm, 58.7℃].

[0131] Every 10 generations, an optimal individual information exchange is carried out between the genetic algorithm population and the particle swarm population. For example, after the 10th generation iteration, the optimal individual in the genetic algorithm population is [995W, 31mTorr, 158sccm, 41sccm, 59℃], with a fitness value of 0.91; the optimal individual in the particle swarm population is [1000W, 32mTorr, 160sccm, 42sccm, 60℃], with a fitness value of 0.93. Replace the individual with the lowest fitness in the genetic algorithm population with the optimal individual in the particle swarm, and similarly replace the particle with the lowest fitness in the particle swarm with the optimal individual in the genetic algorithm.

[0132] The fitness values of two populations are weighted and combined using a dynamic adjustment coefficient to obtain a mixed fitness value. At the initial stage of iteration, the weight of the genetic algorithm is relatively high, set to 0.7, which is beneficial for global search. As the iteration progresses, the weight of the particle swarm gradually increases and reaches 0.7 at the later stage of iteration, which is beneficial for local search. For example, at the 25th generation of iteration, the weight of the genetic algorithm is 0.5 and the weight of the particle swarm is 0.5. At this time, if the fitness value of the optimal individual of the genetic algorithm is 0.95 and the fitness value of the optimal individual of the particle swarm is 0.96, then the mixed fitness value is 0.955.

[0133] Perform local fine-grained search in the neighborhood of the currently obtained optimal solution. For example, the optimal parameter combination at the 40th generation is [1010W, 35mTorr, 165sccm, 45sccm, 65°C]. Set the search step size to 1% of the parameter range. Then, the search step size for the power parameter is 4W, and fine-grained search can be carried out in the range of [1006W, 1014W]. With an interval of 0.5W, evaluate the fitness of 9 candidate points and select the optimal value to replace the current optimal solution.

[0134] Judge the convergence of the mixed fitness value. When the difference between the mixed fitness values of two adjacent generations is less than the preset fitness threshold of 0.001, stop the iteration and output the optimal process parameter combination. For example, the mixed fitness value at the 50th generation is 0.9785, and the mixed fitness value at the 51st generation is 0.9792. The difference is 0.0007, which is less than the threshold of 0.001. At this time, the algorithm converges, and the optimal process parameter combination [1012W, 36mTorr, 168sccm, 47sccm, 67°C] is output.

[0135] In practical applications, this method improves the yield from 92.3% to 94.8% in the optimization of a certain etching process, while reducing the energy consumption by 5.2% and controlling the parameter fluctuation within 3%.

[0136] Figure 4 The following is the experimental data comparison data table for the process parameter optimization method of the embodiment of the present invention:

[0137] This figure shows the performance comparison results of four different optimization algorithms, including a single genetic algorithm, a single particle swarm optimization, the parallel operation of the two algorithms, and the technical solution of the present invention. A comprehensive comparison is carried out from seven dimensions: the improvement in yield, the reduction of process cost, the parameter fluctuation, the convergence speed, the final fitness value, the algorithm stability, and the computational resource occupancy.

[0138] From the specific data, this technical solution performs optimally in multiple key indicators: the yield improvement reaches 5.1%, higher than the other three solutions (3.1%, 3.6% and 4.2% respectively); the process cost reduction reaches 5.2%, significantly better than the other solutions (2.8%, 3.2% and 4.0% respectively); the parameter fluctuation is reduced to 3.7%, which is the lowest among the four solutions (the others are 5.6%, 4.8% and 4.5% respectively); the convergence speed only needs 43 generations to reach the optimal solution, much faster than the other solutions (68, 57 and 52 generations respectively); the final fitness value reaches 0.942, higher than the other solutions (0.862, 0.889 and 0.915 respectively); the algorithm stability (standard deviation) is 0.0068, better than the other solutions (0.0152, 0.0124 and 0.0098 respectively).

[0139] In terms of computing resource occupancy, this technical solution is at a medium-high level, which is a reasonable trade-off to achieve better optimization effects. Through data comparison, it can be seen that this technical solution achieves the best balance between performance and efficiency. Although the computing resource consumption is slightly higher, the improvement in optimization effects brought is significant, especially in the two key economic indicators of yield improvement and cost reduction.

[0140] In the prior art, process parameter optimization mainly uses single optimization algorithms, such as simple genetic algorithms or particle swarm algorithms, and even relies on engineers' experience for manual parameter tuning. These methods have problems such as low optimization efficiency, being prone to falling into local optima, and insufficient ability to explore complex parameter spaces. For example, although the traditional genetic algorithm has strong global search ability, its convergence speed is slow; while the particle swarm algorithm has high local search accuracy, but is prone to premature convergence.

[0141] Starting from the perspective of improving the performance of the optimization algorithm, this application innovatively co-evolves the genetic algorithm and the particle swarm optimization algorithm, making full use of the complementary advantages of the two algorithms. Through the periodic optimal individual information exchange mechanism, the two populations can share the high-quality solutions found, expanding the effective search space; through dynamic adjustment of coefficients, the algorithm gradually shifts from global exploration to local fine search during the search process; a local fine search mechanism is introduced to further improve the accuracy of the optimal solution.

[0142] Compared with the prior art, the co-evolution method of this application has significantly improved in terms of optimization efficiency, solution quality and stability. Experimental data shows that compared with the single genetic algorithm, the convergence speed is increased by about 40%, and the final yield improvement is 0.8 percentage points higher; compared with the single particle swarm algorithm, 90% of the premature convergence situations are avoided, and the quality of the final solution is improved by 12% on average. This improvement is of great significance to the semiconductor manufacturing industry, which can effectively improve the product yield and reduce the manufacturing cost.

[0143] In an alternative embodiment, global optimization of the overall manufacturing process is performed at the process-level optimization layer, and comprehensive evaluation and dynamic adjustment of the optimization results of each optimization layer are carried out based on the Bayesian optimization framework, including:

[0144] Construct a directed acyclic graph of the process flow, where the directed acyclic graph of the process flow includes a set of process nodes, a set of process-dependent edges, and a node weight matrix. Each process node in the set of process nodes contains a set of process parameters, processing time, quality index, and cost factor;

[0145] Based on the directed acyclic graph of the process flow, construct a multi-objective comprehensive evaluation function. The multi-objective comprehensive evaluation function includes a manufacturing cycle term, a product yield term, and a resource utilization term. The manufacturing cycle term is calculated according to the processing time in the process node, the product yield term is calculated according to the quality index, the resource utilization term is calculated according to the cost factor, and weighted combination is performed through dynamic weight coefficients;

[0146] Based on the multi-objective comprehensive evaluation function, establish a Gaussian process regression model, and construct a Bayesian optimization framework using the Matérn kernel function; optimize each optimization layer based on the Bayesian optimization framework to obtain the optimization results of each optimization level. The optimization results include an optimal solution vector, an objective function value, and a confidence level.

[0147] Construct a directed acyclic graph of the process flow. This graph contains a set of process nodes N = {n1, n2,..., n m}, a set of process-dependent edges E = {e1, e2,..., e k}, and a node weight matrix W. Each process node ni contains four key attributes: a set of process parameters Pi = {p1, p2,..., p s}, processing time Ti, quality index Qi, and cost factor Ci. The node weight matrix W is used to represent the association strength between nodes, where the element wij represents the influence degree of node ni on node nj, and the value range is [0, 1]. In practical applications, for example, in the automotive parts manufacturing process, the set of process parameters for the engine block machining process can be set to include a cutting speed of 120 m / min, a feed rate of 0.2 mm / r, a cutting depth of 1.5 mm, etc.; the processing time is 45 minutes; the quality index includes a surface roughness Ra = 1.6 μm and a dimensional accuracy of ±0.01 mm; the cost factor includes equipment depreciation cost of 200 yuan / hour, tool consumption of 30 yuan / piece, labor cost of 80 yuan / hour, etc.

[0148] Construct a multi-objective comprehensive evaluation function \(F\) based on the process flow directed acyclic graph. This evaluation function includes the manufacturing cycle term \(CT\), the product yield rate term \(YR\), and the resource utilization rate term \(RU\), which are weighted and combined through dynamic weight coefficients \(\alpha\), \(\beta\), and \(\gamma\), satisfying \(\alpha+\beta+\gamma = 1\) and \(\alpha,\beta,\gamma\in[0,1]\). The manufacturing cycle term \(CT\) is calculated by the critical path method, that is, find the path with the longest total processing time from the directed acyclic graph, and the sum of the processing times of each node on this path is the manufacturing cycle. The product yield rate term \(YR\) is calculated according to the quality index, and the weighted geometric mean of the quality indexes of each process node is adopted, and the weights are determined according to the importance of the nodes. The resource utilization rate term \(RU\) is calculated according to the cost factors, considering factors such as equipment utilization rate, human resource allocation efficiency, and material utilization rate. In practical applications, the weight coefficients can be dynamically adjusted according to the production strategy of the enterprise. For example, when the delivery date is tight, increase the value of \(\alpha\); when the quality requirements are strict, increase the value of \(\beta\); when the cost control pressure is high, increase the value of \(\gamma\). For example, in the production of precision parts in a certain intelligent manufacturing enterprise, \(\alpha = 0.4\), \(\beta = 0.35\), and \(\gamma = 0.25\) can be set.

[0149] Based on the multi-objective comprehensive evaluation function, a Gaussian process regression model is established. This model uses the Matérn kernel function to construct a Bayesian optimization framework. The hyperparameters of the Matérn kernel function include the length scale parameter \(l\) and the smoothness parameter \(\nu\). In this embodiment, \(l = 0.8\) and \(\nu = 2.5\) are set. The Bayesian optimization framework gradually explores the parameter space through an iterative method. Each iteration includes four steps: First, construct a Gaussian process model according to the existing sample points; second, calculate the acquisition function value, and adopt the expected improvement (EI) strategy; then, select the point with the largest acquisition function value as the next evaluation point; finally, evaluate the objective function value of the new point and update the model.

[0150] For the specific optimization process of each optimization layer, it is divided into three levels: the equipment layer, the process layer, and the control layer. The optimization of the equipment layer mainly focuses on equipment parameters and fixture design. For example, on a CNC machining center, the spindle speed range can be optimized to 2000 - 6000 rpm, the feed speed range to 100 - 500 mm / min, and tool parameters, etc. The optimization of the process layer mainly focuses on the process route and process parameters, including operation sequencing, cutting parameter selection, etc. The optimization of the control layer mainly focuses on production planning and resource scheduling, including job scheduling, personnel allocation, quality inspection strategies, etc. The optimization results of each level include the optimal solution vector X*, the objective function value F*, and the confidence level C. For example, for the precision part machining of a certain intelligent factory, after optimization by this method, the following results are obtained: The optimal solution of the process layer includes a cutting speed of 150 m / min, a feed rate of 0.15 mm / r, and a cutting depth of 1.2 mm, with the corresponding objective function value of 0.82 and a confidence level of 93%; the optimal solution of the equipment layer includes a spindle speed of 4500 rpm and a feed speed of 300 mm / min, with the corresponding objective function value of 0.78 and a confidence level of 91%; the optimal solution of the control layer includes a batch size of 30 pieces and an inspection frequency of inspecting once every 10 pieces, with the corresponding objective function value of 0.75 and a confidence level of 89%.

[0151] After the optimization results of each optimization layer are comprehensively evaluated by the Bayesian optimization framework, dynamic adjustment is carried out. The strategies for dynamic adjustment include parameter fine-tuning, weight reallocation, and constraint condition update. Parameter fine-tuning is achieved through a local search algorithm, which conducts a fine exploration near the optimal solution. Weight reallocation dynamically adjusts the values of α, β, and γ according to the real-time production situation. Constraint condition update updates the constraint conditions of the optimization problem in a timely manner according to changes in production capacity and market demand. Taking the manufacturing of a certain aviation part as an example, during the optimization process, it is found that a certain parameter (cutting speed) in the process layer has a significant impact on product quality. Through dynamic adjustment, the value of β is increased from 0.35 to 0.45, and the value of α is decreased from 0.4 to 0.3. After re-optimization, the cutting speed is adjusted from 150 m / min to 135 m / min, and the product qualification rate is increased from 98.2% to 99.5%.

[0152] By globally optimizing the overall manufacturing process through the process-level optimization layer and comprehensively evaluating and dynamically adjusting the optimization results of each optimization layer based on the Bayesian optimization framework, the overall performance of the manufacturing system can be effectively improved. In the application example of a high-end equipment manufacturing enterprise, after adopting this method, the manufacturing cycle is shortened by 23%, the product yield is increased by 5.8%, the resource utilization rate is increased by 12.7%, and the comprehensive benefits are significantly improved.

[0153] Figure 5 The following is the performance comparison data table for the global optimization of the manufacturing process in the embodiments of the present invention:

[0154] This figure shows the performance comparison data of four optimization methods (traditional sequential optimization, local parameter optimization, graph theory-based optimization and this technical solution), and comprehensively displays various performance indicators and technical characteristics through two tables.

[0155] In terms of performance metrics, this technical solution performed the best: manufacturing cycle time was shortened by 12.4%, significantly exceeding the other solutions (5.7%, 7.3%, and 8.9%, respectively); product yield increased by 7.9%, surpassing the other solutions (3.2%, 4.1%, and 5.6%, respectively); and resource utilization increased by 10.5%, surpassing the other solutions (4.5%, 6.2%, and 7.8%, respectively). In terms of optimization convergence times, this solution reached the optimal solution in just 17 attempts, significantly outperforming the other solutions (38, 31, and 25 attempts, respectively). The multi-objective evaluation function value reached 0.895, with a confidence interval of ±0.018, both demonstrating the best optimization results. Although the computational complexity is O(n²log n), the same as the graph-theory-based optimization solution, the overall performance is significantly superior.

[0156] In terms of technical features, this solution utilizes a multi-layered, adaptive optimization hierarchy, which is more flexible than the single-, dual-, and triple-layer structures of other solutions. While the number of nodes (24) and edges (32) is the same as other solutions, it is unique in that it fully considers dependencies and implements a dynamic weight adjustment mechanism, along with confidence assessment capabilities. While the runtime is longer (85.2 minutes), compared to other solutions (42.5, 58.3, and 76.8 minutes, respectively), this investment of time yields superior optimization results, demonstrating this solution's emphasis on optimizing quality.

[0157] These data fully demonstrate that this technical solution has obvious advantages in optimization effect, stability and adaptability. Although it has higher computing resource consumption and running time, it achieves better overall performance through more complex optimization strategies.

[0158] In an optional embodiment, the method further includes:

[0159] Constructing a combined evaluation index to evaluate the optimization results, establishing an adaptive exploration mechanism based on the evaluation results, and dynamically selecting sampling points through an exploration-utilization balance factor, wherein the exploration-utilization balance factor is adaptively adjusted as the optimization process progresses;

[0160] Real-time monitoring of process parameters during wafer manufacturing. When the drift of the process parameters relative to a preset reference value exceeds a preset drift threshold, a compensation vector is generated based on a proportional-integral-differential controller.

[0161] Apply the compensation vector to the current process parameters to obtain a compensated combination of process parameters, and input the compensated combination of process parameters into the multi-objective comprehensive evaluation function;

[0162] When the output value of the multi-objective comprehensive evaluation function meets the convergence condition and the spatial distribution of the compensated combination of process parameters meets the aggregation condition, determine the final optimization plan.

[0163] To comprehensively evaluate the optimization results, a combined evaluation index is constructed in this embodiment, which comprehensively considers multiple key performance parameters of wafer manufacturing. Specifically, the combined evaluation index is weighted and composed of five sub-indices: wafer surface uniformity, wafer thickness consistency, pattern size accuracy, defect density, and production capacity.

[0164] The weights assigned to each sub-index are adjusted according to the requirements of different product types and process stages. For example, for logic chips at the 14nm process node, the weight of wafer surface uniformity is 0.30, the weight of wafer thickness consistency is 0.25, the weight of pattern size accuracy is 0.25, the weight of defect density is 0.15, and the weight of production capacity is 0.05. For power devices, the weight of wafer thickness consistency may be increased to 0.35, and the other weights are adjusted accordingly.

[0165] Based on the evaluation results, an adaptive exploration mechanism is established, and sampling points are dynamically selected through an exploration-exploitation balance factor. The initial value of this balance factor α is set to 0.7, indicating that more attention is paid to the extensive exploration of the parameter space in the early stage of optimization. As the optimization process progresses, the balance factor will be adaptively adjusted, and its adjustment strategy is as follows:

[0166] When the improvement amplitude of the evaluation index is less than 1% for 5 consecutive iterations, the value of α is decreased by 0.1 to increase the exploitation ratio; when a local optimal trap is found (the evaluation index fluctuates within ±0.5% for 10 consecutive iterations), the value of α is increased by 0.15 to promote the exploration of new regions; when the evaluation index shows a significant improvement (the improvement in a single iteration exceeds 5%), the current value of α remains unchanged.

[0167] The adjustment range of the balance factor is restricted between [0.2, 0.8] to ensure that both exploration and exploitation strategies can be applied to a certain extent during the optimization process. In an actual production environment, this mechanism enables the system to quickly adapt and find a better parameter combination under changing process conditions.

[0168] By real-time monitoring of key process parameters during wafer manufacturing, process stability is ensured. The monitored parameters include exposure dose, focus deviation, development time, etching power, and temperature, etc. Each parameter is set with a preset reference value and an allowable drift threshold. For example, the preset reference value of the exposure dose is 30 mJ / cm², and the allowable drift threshold is ±2%; the preset reference value of the focus deviation is 0 μm, and the allowable drift threshold is ±0.15 μm; the preset reference value of the development time is 60 seconds, and the allowable drift threshold is ±3 seconds.

[0169] When the drift amount of any process parameter exceeds the preset drift threshold, the system generates a compensation vector based on a proportional-integral-derivative controller. The three key parameters of this controller are configured as follows: the proportional coefficient Kp is set to 0.65, which is used to provide an immediate response proportional to the deviation; the integral coefficient Ki is set to 0.15, which is used to eliminate long-term steady-state errors; the derivative coefficient Kd is set to 0.20, which is used to predict system behavior and improve response characteristics.

[0170] The generation of the compensation vector takes into account the magnitude, duration, and trend of the parameter drift. For example, when it is detected that the exposure dose has been 3% higher for 5 minutes, the compensation vector generated by the system will include an exposure dose adjustment of -2.1%; when the development time fluctuates within the range of ±4 seconds and there is no obvious trend, the system will calculate the compensation amount based on the mean deviation in the most recent 30 minutes.

[0171] After the compensation vector is generated, it immediately acts on the current process parameters to obtain the compensated process parameter combination. During the parameter compensation process, the system ensures that the adjustment amplitude of each parameter does not exceed the safety threshold. For example, the single adjustment of the exposure dose does not exceed 5%, and the temperature adjustment does not exceed 2°C, to prevent process instability caused by overcorrection.

[0172] The compensated process parameter combination is input into a multi-objective comprehensive evaluation function, which comprehensively considers three dimensions: product yield, production efficiency, and cost. This function uses the weighted summation method and assigns different weights to different objectives. In this implementation, the yield weight is 0.6, the production efficiency weight is 0.3, and the cost weight is 0.1.

[0173] The output value of the multi-objective comprehensive evaluation function is used to determine whether the optimization process reaches the convergence condition. The convergence condition is set as follows: the relative change rate of the output value of the evaluation function is less than 0.8% in 15 consecutive iterations, or the preset maximum number of iterations of 300 times is reached.

[0174] The system evaluates whether the spatial distribution of the compensated process parameter combination meets the aggregation condition. The aggregation condition is defined as: 95% of the candidate points in the parameter space are located within 20% of the parameter space, and the Euclidean distance between the best point and the second-best point is less than the preset threshold (5% of the diameter of the parameter space).

[0175] When the convergence condition and the aggregation condition are both satisfied, the system determines the final optimization plan. The final optimization plan is determined by the parameter combination with the highest current evaluation function value, and at the same time, the sub-optimal plan is recorded as an alternative. For example, in a certain optimization, the determined final plan is: exposure dose 31.2 mJ / cm², focus deviation -0.05 μm, development time 58 seconds, etching power 850 W, temperature 135 °C, and the output value of the comprehensive evaluation function of this plan reaches 0.876 (with a full score of 1).

[0176] It was verified on a 28nm process production line in a certain 12-inch wafer fab. Before optimization, the critical dimension uniformity (CDU) of this production line was 8.2 nm, the yield was 91.3%, and the single-wafer processing time was 58 minutes. After applying this method for process parameter optimization, the CDU was improved to 6.4 nm, the yield was increased to 94.7%, the single-wafer processing time was shortened to 52 minutes, and the comprehensive production efficiency was increased by 11.3%.

[0177] In response to sudden process fluctuations, the traditional method needs to suspend production and spend an average of 4 hours for process debugging, while this method can respond in real time during the production process, and the parameter adjustment time is shortened to an average of 22 minutes, greatly reducing the downtime and material losses.

[0178] Through six months of continuous application, this method has shown good adaptability and stability under different product types and process conditions, demonstrating its practical value and effectiveness in the actual wafer manufacturing environment.

[0179] In the second aspect of the embodiments of the present invention,

[0180] A wafer-level test yield prediction and process optimization system based on big data is provided, including:

[0181] The first unit is used to obtain process parameter data, equipment status data, defect detection data, and test yield data during the wafer manufacturing process;

[0182] The second unit is used to construct a time-series sensitive multi-modal data fusion network. The multi-modal data fusion network adopts a hybrid architecture combining the long short-term memory network LSTM and the graph neural network GNN. It captures the time-series characteristics of process parameter data, equipment status data, and defect detection data through LSTM, uses GNN to establish the spatial topological relationship of test yield data, and adaptively assigns weights to the spatial features corresponding to the time-series characteristics and the spatial topological relationship through an attention mechanism to predict the wafer test yield and obtain the predicted yield value;

[0183] The third unit is used to trigger the process optimization process when the predicted yield value is lower than the preset yield threshold;

[0184] The fourth unit is used to construct a hierarchical progressive process optimization system. The hierarchical progressive process optimization system includes an equipment parameter optimization layer and a process-level optimization layer. The equipment parameter optimization layer performs local optimization on a single process parameter and population optimization on the parameter combinations of multiple related devices through the co-evolution method of the genetic algorithm and the particle swarm optimization algorithm. The process-level optimization layer performs global optimization on the overall manufacturing process, and comprehensively evaluates and dynamically adjusts the optimization results of each optimization layer based on the Bayesian optimization framework;

[0185] The fifth unit is used to adjust process parameters according to the optimization scheme generated by the hierarchical progressive process optimization system and generate process optimization instructions.

[0186] In the third aspect of the embodiments of the present invention,

[0187] There is provided an electronic device, including:

[0188] A processor;

[0189] A memory for storing instructions executable by the processor;

[0190] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0191] In the fourth aspect of the embodiments of the present invention,

[0192] There is provided a computer-readable storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0193] The present invention can be a method, a device, a system, and / or a computer program product. The computer program product can include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0194] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting wafer - level test yield and optimizing process based on big data, characterized in that, Including: Obtaining process parameter data, equipment status data, defect detection data, and test yield data during the wafer manufacturing process; Constructing a time-series sensitive multi-modal data fusion network, the multi-modal data fusion network adopting a hybrid architecture combining a long short-term memory network (LSTM) and a graph neural network (GNN). Capturing the time-series characteristics of process parameter data, equipment status data, and defect detection data through LSTM, establishing the spatial topological relationship of test yield data using GNN, and adaptively allocating weights to the spatial features corresponding to the time-series characteristics and spatial topological relationship through an attention mechanism to predict the wafer test yield and obtain a predicted yield value; When the predicted yield value is lower than a preset yield threshold, triggering a process optimization process; Constructing a hierarchical progressive process optimization system, the hierarchical progressive process optimization system including an equipment parameter optimization layer and a process-level optimization layer. The equipment parameter optimization layer locally optimizes a single process parameter and globally optimizes the parameter combinations of multiple related equipment through a co-evolution method of a genetic algorithm and a particle swarm optimization algorithm. At the process-level optimization layer, the overall manufacturing process is globally optimized, and the optimization results of each optimization layer are comprehensively evaluated and dynamically adjusted based on a Bayesian optimization framework; Adjusting the process parameters according to the optimization scheme generated by the hierarchical progressive process optimization system to generate a process optimization instruction.

2. The method according to claim 1, wherein The process parameter data includes etching depth, lithography alignment accuracy, metal deposition thickness, and doping concentration; the equipment status data includes chamber pressure, radio frequency power, gas flow rate, and temperature; The defect detection data includes defect size, defect type, defect density, and defect distribution; the test yield data includes functional test yield, parameter test yield, and reliability test yield.

3. The method according to claim 1, wherein Adapting weight allocation to the spatial features corresponding to the time-series characteristics and spatial topological relationship through an attention mechanism to predict the wafer test yield and obtain a predicted yield value includes: Using a long short-term memory network to extract features from the process parameter data, the equipment status data, and the defect detection data, and performing selective memory through a forget gate, an input gate, and an output gate to obtain time-series characteristics; Constructing a graph structure based on the test yield data, taking the bad points on the wafer as the nodes of the graph structure, taking the spatial correlation relationship between adjacent bad points as the edges of the graph structure, and using a graph neural network to extract features from the graph structure and obtaining spatial features through node feature aggregation; Taking the time-series characteristics as query vectors, taking the spatial features as key vectors and value vectors, calculating the similarity between the query vectors and the key vectors to obtain attention weights, and performing weighted summation on the value vectors based on the attention weights to obtain a fused feature; Inputting the fused feature into a prediction layer, and obtaining a predicted yield value through a non-linear transformation. The prediction layer uses a sigmoid activation function to map the prediction output to between 0 and 1.

4. The method according to claim 1, characterized in that, Through a co-evolution method of a genetic algorithm and a particle swarm optimization algorithm, locally optimizing a single process parameter at the parameter-level optimization layer includes: Encode the process parameters to be optimized as real - valued chromosomes to generate a parameter encoding vector. The parameter encoding vector contains n process parameter components, and corresponding upper and lower limit values of the parameters are set for each process parameter component; construct a fitness function with the yield improvement amount, process cost, and parameter fluctuation degree as the optimization objectives, and perform weighted combination on the yield improvement amount, the process cost, and the parameter fluctuation degree through weight coefficients; Generate an initial genetic algorithm population and an initial particle swarm population. The initial genetic algorithm population contains N chromosome individuals, and the initial particle swarm population contains M particle individuals; Perform a selection operation on the genetic algorithm population. Use the roulette wheel selection method to calculate the selection probability according to the individual fitness value, and select high - quality individuals to enter the mating pool; perform an arithmetic crossover operation on the individuals in the mating pool, generate offspring individuals through the crossover coefficient, and perform a Gaussian mutation operation on the offspring individuals; Update the velocity and position information of each particle in the particle swarm, and adjust the particle position according to the updated velocity; Exchange the optimal individual information between the genetic algorithm population and the particle swarm population every preset number of generations. Inject the optimal individual information of the genetic algorithm population into the particle swarm population, and inject the optimal individual information of the particle swarm population into the genetic algorithm population; Use a dynamic adjustment coefficient to perform weighted combination on the fitness of the genetic algorithm population and the particle swarm population to obtain a mixed fitness value; perform local fine - grained search in the neighborhood of the currently obtained optimal solution, and fine - tune the optimal solution through the search step size and the search direction vector; judge the convergence of the mixed fitness value. When the difference between the mixed fitness values of two adjacent generations is less than the preset fitness threshold, output the optimal process parameter combination.

5. The method according to claim 1, wherein Perform global optimization on the overall manufacturing process at the process - level optimization layer, and perform comprehensive evaluation and dynamic adjustment on the optimization results of each optimization layer based on the Bayesian optimization framework, including: Construct a directed acyclic graph of the process flow. The directed acyclic graph of the process flow includes a set of process nodes, a set of process - dependent edges, and a node weight matrix. Each process node in the set of process nodes contains a set of process parameters, processing time, quality index, and cost factor; Construct a multi - objective comprehensive evaluation function based on the directed acyclic graph of the process flow. The multi - objective comprehensive evaluation function includes a manufacturing cycle term, a product yield term, and a resource utilization rate term. Calculate the manufacturing cycle term according to the processing time in the process node, calculate the product yield term according to the quality index, calculate the resource utilization rate term according to the cost factor, and perform weighted combination through dynamic weight coefficients; Establish a Gaussian process regression model based on the multi - objective comprehensive evaluation function, and construct a Bayesian optimization framework using the Matérn kernel function; perform optimization on each optimization layer based on the Bayesian optimization framework to obtain the optimization results of each optimization level. The optimization results include an optimal solution vector, an objective function value, and a confidence level.

6. The method according to claim 5, characterized in that, The method further includes: Construct a combined evaluation index to evaluate the optimization result, establish an adaptive exploration mechanism according to the evaluation result, and dynamically select sampling points through an exploration-exploitation balance factor, where the exploration-exploitation balance factor is adaptively adjusted as the optimization process progresses; Monitor the process parameters in the wafer manufacturing process in real time. When the drift amount of the process parameters relative to the preset reference value exceeds the preset drift threshold, generate a compensation vector based on a proportional-integral-derivative controller; Apply the compensation vector to the current process parameters to obtain a compensated process parameter combination, and input the compensated process parameter combination into the multi-objective comprehensive evaluation function; When the output value of the multi-objective comprehensive evaluation function meets the convergence condition and the spatial distribution of the compensated process parameter combination meets the aggregation condition, determine the final optimization plan.

7. A wafer-level test yield prediction and process optimization system based on big data, for implementing the method according to any one of the preceding claims 1-6, characterized in that, It includes: The first unit is used to obtain process parameter data, equipment status data, defect detection data, and test yield data in the wafer manufacturing process; The second unit is used to construct a time-series sensitive multi-modal data fusion network. The multi-modal data fusion network adopts a hybrid architecture combining a long short-term memory network (LSTM) and a graph neural network (GNN). The LSTM captures the time-series characteristics of process parameter data, equipment status data, and defect detection data, and the GNN establishes the spatial topological relationship of test yield data. An attention mechanism is used to adaptively allocate weights to the spatial features corresponding to the time-series characteristics and the spatial topological relationship to predict the wafer test yield and obtain a predicted yield value; The third unit is used to trigger the process optimization process when the predicted yield value is lower than the preset yield threshold; The fourth unit is used to construct a hierarchical progressive process optimization system. The hierarchical progressive process optimization system includes an equipment parameter optimization layer and a process-level optimization layer. The equipment parameter optimization layer locally optimizes a single process parameter and globally optimizes the parameter combinations of multiple related equipment through a co-evolution method of a genetic algorithm and a particle swarm optimization algorithm. The overall manufacturing process is globally optimized at the process-level optimization layer, and the optimization results of each optimization layer are comprehensively evaluated and dynamically adjusted based on a Bayesian optimization framework; The fifth unit is used to adjust the process parameters according to the optimization plan generated by the hierarchical progressive process optimization system and generate a process optimization instruction.

8. An electronic device, characterized in that, It includes: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 6 is implemented.

Images