A method and apparatus for generating synthetic data

Abstract

Embodiments of the present specification relate to a method and device for generating synthetic data, the method comprising: obtaining at least one target synthetic material used for generating synthetic data, the types of the target synthetic material including real data collected by sensors of equipment in a semiconductor manufacturing process, data rules and / or business rules related to semiconductor manufacturing, and a preset simulation data pattern; based on the types of the target synthetic material, processing the target synthetic material using a corresponding data synthesis model to obtain corresponding first synthetic data, the data volume of the first synthetic data being determined according to a preset first threshold; the data synthesis model includes an artificial intelligence data synthesis model corresponding to the real data, a rule data synthesis model corresponding to the data rules and / or business rules, and a simulation data synthesis model corresponding to the simulation data pattern.

Description

Technical Field

[0001] This specification relates to one or more embodiments in the field of semiconductor manufacturing, and more particularly to a method and apparatus for generating synthetic data. Background Technology

[0002] Semiconductor manufacturing, as one of the most complex processes in modern industry, involves hundreds or even thousands of intricate steps, including photolithography, etching, thin film deposition, diffusion, and cleaning. With the increasing functionality and integration of integrated circuit chips, the requirements for product design, equipment maintenance, process control, and quality management are becoming increasingly stringent. Building virtual factories using digital twins, artificial intelligence, and intelligent equipment to recreate the factory in virtual reality and achieve efficient, unmanned, and continuously improving precise control has become urgent. A virtual factory can be viewed as an industrial collaborative environment, focusing on simulating the actual production data flow and capacity of a factory, reflecting the real-world factory conditions. A virtual factory should be able to generate data at different granularities, and these data sources can, in turn, be used to develop, test, and evaluate various data analytics applications.

[0003] However, the industry faces numerous challenges in building data analytics applications using real-world data. These data limitations significantly hinder the wider application and deployment of artificial intelligence. Among these, obtaining a sufficient amount of real-world data for algorithm development and model validation is a primary issue in application implementation. However, the data types, sample sizes, data distributions, and outlier occurrences vary across different process stages, leading to diverse data requirements. Therefore, a flexible and universally applicable method for generating synthetic data is needed to continuously improve its quality and ensure its consistency, reliability, and diversity with real-world data. Summary of the Invention

[0004] This specification describes one or more embodiments of a method and apparatus for generating synthetic data, which aims to automatically generate synthetic data in batches based on real data, data rules, business rules, and other information generated during the semiconductor manufacturing process.

[0005] Firstly, a method for generating synthetic data is provided, including:

[0006] Acquire at least one target synthetic material used to generate synthetic data. The types of target synthetic materials include real data collected by sensors of equipment during semiconductor manufacturing, data rules and / or business rules related to semiconductor manufacturing, and preset simulation data patterns.

[0007] Based on the type of the target synthetic material, the target synthetic material is processed using the corresponding data synthesis model to obtain the corresponding first synthetic data. The amount of data in the first synthetic data is determined according to a preset first threshold. The data synthesis model includes an artificial intelligence data synthesis model corresponding to the real data, a rule data synthesis model corresponding to the data rules and / or business rules, and a simulated data synthesis model corresponding to the simulated data mode.

[0008] In one possible implementation, the method further includes:

[0009] Based on the specific purpose of the first synthetic data, the weight values ​​of each quality assessment indicator are determined; the quality assessment indicators include: business logic, data coverage, data similarity, correlation between data, and machine learning effectiveness.

[0010] The first synthetic data is scored based on each quality assessment index, and the scores are weighted and summed according to their corresponding weight values ​​to obtain the quality assessment score.

[0011] When the quality assessment score is greater than a preset score threshold, the target synthetic material is processed using a corresponding data synthesis model based on the type of the target synthetic material to obtain corresponding second synthetic data. The data volume of the second synthetic data is determined according to a preset second threshold; the data volume of the second synthetic data is greater than the data volume of the first synthetic data.

[0012] In one possible implementation, the target synthetic material includes real data, and the data synthesis model includes an artificial intelligence data synthesis model; based on the type of the target synthetic material, the target synthetic material is processed using the corresponding data synthesis model to obtain corresponding first synthetic data, including:

[0013] At least the real data is used to train each candidate model in the artificial intelligence data synthesis model, and based on the performance of each trained candidate model, one or more candidate models are determined as the first target model.

[0014] The first target model is optimized using a preset hyperparameter optimization algorithm to obtain the second target model;

[0015] The second target model is used to generate the first synthetic data.

[0016] In one possible implementation, before training the candidate models in the artificial intelligence data synthesis model using the real data, the method further includes:

[0017] Data analysis is performed on the real data to obtain its data characteristics, which include at least one of the following: statistical characteristics, data type of each field, data structure of each field, and logical analysis characteristics.

[0018] Send the data features to the user operation platform;

[0019] The system receives user-configured data rule constraints and / or business logic constraints for the synthesized data from the user operation platform. The data rule constraints include at least one of the following: data type, data distribution, data range, and primary / foreign key relationships. The business logic constraints include at least one of the following: the subordinate relationship between multiple fields in semiconductor manufacturing, and the constraint of a specific semiconductor physical mechanism model that one or more fields are subject to.

[0020] In one possible implementation, training each candidate model in the artificial intelligence data synthesis model using at least the real data includes:

[0021] Using the real data, data rule constraints, and / or business logic constraints, train each candidate model in the artificial intelligence data synthesis model.

[0022] In one possible implementation, before training the candidate models in the artificial intelligence data synthesis model using the real data, the method further includes:

[0023] The real data is cleaned to remove or repair any unreasonable data; the unreasonable data includes at least one of the following: missing values, outliers, and duplicate values.

[0024] In one possible implementation, the target composite material includes data rules and / or business rules, and the data synthesis model includes a rule-based data synthesis model; based on the type of the target composite material, the target composite material is processed using the corresponding data synthesis model to obtain corresponding first composite data, including:

[0025] Receive user-configured data distribution type, shape factor, and size factor from the user operation platform;

[0026] Based on the data rules and / or business rules, as well as the data distribution type, shape factor, and size factor, the rule-based data synthesis model is used to sample from the data distribution type, and perturbation noise data is added to the sampling results to obtain the first synthesized data.

[0027] In one possible implementation, the target composite material includes a simulated data pattern, and the data compositing model includes a simulated data compositing model; based on the type of the target composite material, the target composite material is processed using the corresponding data compositing model to obtain corresponding first composite data, including:

[0028] The system receives one or more target waveforms selected by the user from the simulation data mode from the user operation platform; the target waveforms include at least: sine wave, square wave, triangle wave, and sawtooth wave.

[0029] The simulated data synthesis model is used to superimpose the one or more target waveforms to obtain a target simulated waveform, and sampling is performed from the target simulated waveform to obtain the first synthesized data.

[0030] Secondly, an apparatus for generating synthetic data is provided, comprising:

[0031] The acquisition unit is configured to acquire at least one target synthetic material used to generate synthetic data. The types of target synthetic materials include real data collected by sensors of equipment during semiconductor manufacturing, data rules and / or business rules related to semiconductor manufacturing, and preset simulation data modes.

[0032] The first synthesis unit is configured to process the target synthetic material using a corresponding data synthesis model based on the type of the target synthetic material to obtain corresponding first synthetic data. The amount of data in the first synthetic data is determined according to a preset first threshold. The data synthesis model includes an artificial intelligence data synthesis model corresponding to the real data, a rule data synthesis model corresponding to the data rules and / or business rules, and a simulated data synthesis model corresponding to the simulated data mode.

[0033] In one possible implementation, the device further includes:

[0034] The weight determination unit is configured to determine the weight values ​​of each quality assessment indicator based on the specific purpose of the first synthetic data; the quality assessment indicators include: business logic, data coverage, data similarity, correlation between data, and machine learning effectiveness.

[0035] The quality assessment framework is configured to score the first synthetic data based on each quality assessment indicator, and then sum the scores according to their corresponding weight values ​​to obtain a quality assessment score.

[0036] The second synthesis unit is configured to, when the quality assessment score is greater than a preset score threshold, process the target synthesis material using a corresponding data synthesis model based on the type of the target synthesis material to obtain corresponding second synthesis data, wherein the data volume of the second synthesis data is determined according to a preset second threshold; the data volume of the second synthesis data is greater than the data volume of the first synthesis data.

[0037] Thirdly, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of the first aspect.

[0038] Fourthly, a computing device is provided, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, it implements the method of the first aspect.

[0039] The embodiments of this specification propose a method and apparatus for generating synthetic data. Based on the needs of the actual scenario, the method adaptively selects the best way to generate synthetic data and automatically generates synthetic data in batches based on real data, data rules, business rules and other information generated in the semiconductor manufacturing process. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the various embodiments disclosed in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only a few embodiments disclosed in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 A schematic diagram illustrating an implementation scenario of a method for generating synthetic data according to one embodiment is shown.

[0042] Figure 2 A flowchart illustrating a method for generating synthetic data according to one embodiment is shown;

[0043] Figure 3 A schematic diagram illustrating a synthetic data evaluation metric according to one embodiment is shown;

[0044] Figure 4 A schematic diagram illustrating a neural network architecture search according to one embodiment is shown;

[0045] Figure 5 A flowchart illustrating a method for generating synthetic data based on real data according to one embodiment is shown.

[0046] Figure 6 A schematic block diagram of an apparatus for generating synthetic data according to one embodiment is shown. Detailed Implementation

[0047] The solution provided in this specification will now be described with reference to the accompanying drawings.

[0048] As mentioned earlier, the industry faces numerous challenges when building data analytics applications using real-world data: semiconductor equipment manufacturers sell different brands and models of process equipment to wafer fabs, resulting in vastly different transmission protocols, parsing methods, and parameter categories, leading to high data acquisition costs; data silos between machines create information islands, hindering effective data exchange and sharing; insufficient data diversity makes it difficult to cover all anomalies, or specific application scenarios are difficult to reproduce in the real world; inconsistent data collection frequencies and non-standard data formats among equipment deployed by different suppliers at different times make it difficult to guarantee data quality; process equipment data simulation, simulation software systems, and key process parameters are considered core secrets by various companies, posing information security challenges and restrictions on data acquisition, use, and sharing, etc. These data limitations significantly hinder the wider application and deployment of artificial intelligence. Obtaining real, sufficient sample data for algorithm development and model validation is the primary issue in the application implementation process.

[0049] Therefore, the acquisition of accurate and reliable data is crucial throughout the entire value chain of integrated circuit manufacturing. However, the data types, sample sizes, data distributions, and outlier occurrences vary significantly across different process stages, leading to diverse data requirements. Consequently, a flexible and universally applicable automated data synthesis method for virtual factories is needed to continuously improve the quality of synthesized data and ensure its consistency, reliability, and diversity with real-world data. Furthermore, refined and high-fidelity synthesized data can be used to optimize the modeling and analysis applications across the entire data chain lifecycle, including real-world data models, business models, geometric models, and simulation models, continuously driving innovation and promoting the application and deployment of artificial intelligence in integrated circuit manufacturing.

[0050] Conceptually, synthetic data refers to self-labeled data generated by computer simulation techniques or algorithms. It can mathematically or statistically reflect the properties of real-world data, thus serving as a substitute for real-world data to train, test, and validate algorithm models. Thanks to the continuous emergence of artificial intelligence algorithms in recent years, the types of synthetic data have expanded, and their quality has continuously improved. However, research on data synthesis technology is still in its early stages. Applications of data synthesis technology in the semiconductor manufacturing industry are rarely reported, and the current research does not cover the entire process of building and applying synthetic data models. Specifically, it cannot adaptively select the best synthetic data algorithm based on the actual scenario and optimize the hyperparameters of the synthetic data model. In other words, it cannot use automated data-driven methods to generate optimal synthetic data based on real-world scenarios.

[0051] To solve the above problems, Figure 1 A schematic diagram illustrating an implementation scenario of a method for generating synthetic data according to one embodiment is shown. Figure 1 As shown, firstly, the synthetic materials used to generate the synthetic data are acquired. These materials can be mainly divided into three categories: real data collected by sensors in semiconductor manufacturing equipment, data rules and / or business rules related to semiconductor manufacturing, and preset simulation data patterns. Then, based on the type of synthetic material, the corresponding data synthesis model is used to process the material, obtaining and outputting the corresponding type of synthetic data. This synthetic data includes synthetic data generated based on artificial intelligence models, rule-based synthetic data, and simulated data generated based on simulation data patterns. The synthetic data can be used in subsequent semiconductor-related data modeling and other business processes. Data synthesis models include artificial intelligence data synthesis models, rule-based data synthesis models, and simulation data synthesis models, corresponding to the three types of synthetic materials mentioned above. Furthermore, when generating synthetic data using real data, real data can be mixed with synthetic data to obtain output synthetic data with richer quantity and content.

[0052] The following describes the specific implementation steps of the method for generating synthetic data described above, with reference to specific embodiments. Figure 2 A flowchart illustrating a method for generating synthetic data according to one embodiment is provided. The entity executing the method can be any platform, server, or device cluster with computing and processing capabilities. Figure 2 As shown, the method includes at least the following steps: Step 202, acquiring at least one target synthetic material used to generate synthetic data, wherein the types of target synthetic materials include real data collected by sensors of equipment during semiconductor manufacturing, data rules and / or business rules related to semiconductor manufacturing, and preset simulated data patterns; Step 203, processing the target synthetic material using a corresponding data synthesis model based on the type of target synthetic material to obtain corresponding first synthetic data, wherein the data volume of the first synthetic data is determined according to a preset first threshold; the data synthesis model includes an artificial intelligence data synthesis model corresponding to the real data, a rule data synthesis model corresponding to the data rules and / or business rules, and a simulated data synthesis model corresponding to the simulated data pattern. The specific execution process of each of the above steps is described below.

[0053] First, in step 202, at least one target synthetic material used to generate synthetic data is obtained. The types of target synthetic materials include real data collected by sensors of equipment during semiconductor manufacturing, data rules and / or business rules related to semiconductor manufacturing, and preset simulation data modes.

[0054] In one embodiment, when multiple target composite materials are acquired simultaneously, each target composite material has a priority order for use, including that real data has a higher priority than data rules and / or business rules, and data rules and / or business rules have a higher priority than simulated data patterns.

[0055] When real data collected by semiconductor device sensors can be obtained, the synthetic data generated based on the real data is often the closest to the real data, and therefore has the best authenticity. When real data cannot be obtained, data rules and / or business rules in the semiconductor manufacturing process can be collected. Although they are not as direct as real data, they are still directly related to the real semiconductor manufacturing process, so the synthetic data generated based on data rules and / or business rules has a slightly lower authenticity. When neither of the above two types of synthetic materials can be obtained, the synthetic data can only be generated based on a preset data pattern, past experience, and the purpose of the synthetic data in this case, simulating the possible distribution of real data.

[0056] Then, in step 203, based on the type of the target synthetic material, the target synthetic material is processed using the corresponding data synthesis model to obtain the corresponding first synthetic data. The amount of data in the first synthetic data is determined according to a preset first threshold. The data synthesis model includes an artificial intelligence data synthesis model corresponding to the real data, a rule data synthesis model corresponding to the data rules and / or business rules, and a simulated data synthesis model corresponding to the simulated data mode.

[0057] The first synthesized data can be directly used as the final synthesized data for output. Alternatively, it can be used as test data to evaluate the performance of the corresponding data synthesis model. When the model performance meets the preset requirements, the data synthesis model is then used to generate more synthesized data as the final synthesized data for output.

[0058] Based on this, in some possible implementations, the method further includes steps 201, 204 and 205.

[0059] In step 201, based on the specific purpose of the first synthesized data, the weight values ​​of each quality assessment indicator are determined; the quality assessment indicators include: business logic, data coverage, data similarity, data correlations, and machine learning efficiency.

[0060] The weights of each quality assessment metric can be determined based on the specific purpose of the synthesized data. For example, if the synthesized data is to be used for software testing, the weights would likely be ranked as follows: business logic > data coverage > data similarity > data correlation > machine learning effectiveness. More specifically, the weights could be set as follows: business logic 0.5, data coverage 0.3, data similarity 0.15, data correlation 0.05, and machine learning effectiveness 0. Conversely, if the synthesized data is to be used to train predictive algorithms, the weights would likely be ranked as follows: machine learning effectiveness > data correlation > business logic > data coverage > data similarity. More specifically, the weights could be set as follows: machine learning effectiveness 0.35, data correlation 0.25, business logic 0.2, data coverage 0.15, and data similarity 0.05.

[0061] Specifically, business logic scores can be evaluated based on business rules and business schemes; data coverage scores can be evaluated based on category coverage, boundary adhesion, statistical similarity, and range coverage; data similarity scores can be evaluated based on Bayesian network similarity (BNLikelihood), Bayesian network log similarity (BNLogLikelihood), and Gaussian mixture similarity (GMLikelihood); data correlation scores can be evaluated based on total variation complement (TVComplement), contingency similarity, correlation similarity, and Kolmogorov-Smirnov complement (KSComplement); and machine learning effectiveness scores can be evaluated based on time series, binary classification, multiclass classification, and regression. Various synthetic data quality assessment metrics can be used as follows: Figure 3 As shown. Figure 3 A schematic diagram of a synthetic data evaluation metric according to one embodiment is shown.

[0062] Then, in step 204, the first synthetic data is scored based on each quality assessment index, and the scores are weighted and summed according to their corresponding weight values ​​to obtain the quality assessment score.

[0063] In step 205, when the quality assessment score is greater than a preset score threshold, the target synthetic material is processed using a corresponding data synthesis model based on the type of the target synthetic material to obtain corresponding second synthetic data. The data volume of the second synthetic data is determined according to a preset second threshold; the data volume of the second synthetic data is greater than the data volume of the first synthetic data.

[0064] Steps 201, 204, and 205 describe how a small amount of first synthetic data generated by the model is used as test data for quality evaluation. When the quality evaluation score is greater than a preset score threshold, a larger amount of second synthetic data is generated and output as the final synthetic data. The number of first and second synthetic data generated by the model is constrained by a first threshold and a second threshold, respectively.

[0065] The above describes the overall process of generating synthetic data based on synthetic materials and evaluating the quality of the synthetic data. The following will describe, based on several embodiments, the specific methods for generating synthetic data in step 203 using different data synthesis models for different types of target synthetic materials.

[0066] In one embodiment, the target synthetic material includes real data, and the data synthesis model includes an artificial intelligence data synthesis model. In this case, step 203 specifically includes: step (a) training each candidate model in the artificial intelligence data synthesis model using at least the real data, and determining one or more candidate models as a first target model based on the performance of each trained candidate model; step (b) optimizing the hyperparameters of the first target model using a preset hyperparameter optimization algorithm to obtain a second target model; and step (c) generating first synthetic data using the second target model.

[0067] Specifically, step (a) iterates by training different models using the same input data, selecting the best-performing model. Furthermore, techniques such as fusion and cascading can be used to integrate multiple models for better results. Candidate models include at least: Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), Gaussian Copula, Bayesian Methods, Transformer models, Diffusion Models, and Neural Radiance Fields (NeRF). Model evaluation methods used to assess the performance of each trained candidate model include at least: Early Stop, Low Fidelity, Surrogate, and Parameter Sharing.

[0068] The hyperparameter optimization algorithms in step (b) include at least: Random Search (RS), Grid Search (GS), Gradient Descent (GD), Hyperband algorithm, Reinforcement Learning (RL), Evolutionary Algorithms (EA), Bayesian Optimization (BO), and Bayesian Optimization and Hyperband (BOHB).

[0069] The steps of model selection, model evaluation, and hyperparameter optimization based on steps (a) and (b) can be collectively referred to as Neural Architecture Search (NAS). A schematic diagram of NAS can be shown as follows: Figure 4 As shown. Figure 4 A schematic diagram illustrating a neural network architecture search according to one embodiment is shown.

[0070] In a more specific embodiment, before step (a), step 203 further includes: performing data analysis on the real data to obtain its data characteristics, the data characteristics including at least one of the following: statistical characteristics, data types of each field, data structures of each field, and logical analysis characteristics; then, sending the data characteristics to the user operation platform; next, receiving user-configured data rule constraints and / or business logic constraints for the synthesized data from the user operation platform; the data rule constraints including at least one of the following: data types, data distribution, data range, primary and foreign key relationships; the business logic constraints including at least one of the following: subordinate relationships between multiple fields in semiconductor manufacturing, and constraints of a specific semiconductor physical mechanism model that one or more fields obey.

[0071] Specifically, the statistical characteristics include at least one of the following: minimum value, maximum value, median value, mean, variance, skewness, and kurtosis. The data types of each field include at least one of the following: timestamp, integer, floating-point, nominal, boolean, and string. The data structure and logical analysis characteristics of each field include at least: primary / foreign key relationships in the data table, and correlations between fields. The hierarchical relationships between the multiple fields in semiconductor manufacturing can include: hierarchical relationships between fields such as Die ID, Wafer ID, Lot ID, and Product ID in the semiconductor manufacturing process.

[0072] Based on the data rule constraints and / or business logic constraints received in the foregoing embodiments, the trained artificial intelligence model can be further constrained. In this case, step (a) specifically includes: using the real data and the data rule constraints and / or business logic constraints to train each candidate model in the artificial intelligence data synthesis model.

[0073] In another, more specific embodiment, before step (a), step 203 further includes: performing data cleaning on the real data, removing or repairing unreasonable data therein; the unreasonable data includes at least one of the following: missing values, outliers, and duplicate values.

[0074] Furthermore, when the target composite material includes real data, in step 202, when obtaining the target composite material containing real data, it can be done by receiving one or more files uploaded by the user operation platform, in which the real data is recorded; or it can be done by connecting to a database that stores the corresponding real data and then obtaining the relevant real data from the database. There is no limitation here.

[0075] In one embodiment, after generating synthetic data using real data in step 205, the method further includes: mixing the real data with the second synthetic data to obtain third synthetic data.

[0076] Based on the foregoing embodiments, the following can be obtained: Figure 5 The flowchart shown. Figure 5 A flowchart illustrating a method for generating synthetic data based on real data according to one embodiment is shown.

[0077] The above describes the specific method and process for generating synthetic data based on real data in step 203.

[0078] In one embodiment, the target synthetic material includes data rules and / or business rules, and the data synthesis model includes a rule-based data synthesis model. In this case, step 203 specifically includes: receiving user-configured data distribution type, shape factor, and size factor from the user operation platform. Then, based on the data rules and / or business rules, as well as the data distribution type, shape factor, and size factor, sampling is performed from the data distribution type using the rule-based data synthesis model, and perturbation noise data is added to the sampling results to obtain the first synthetic data.

[0079] The data distribution type must include at least one of the following: Gaussian distribution, exponential distribution, Poisson distribution, Weibull distribution, and log-normal distribution.

[0080] In one embodiment, the target synthetic material includes a simulated data pattern, and the data synthesis model includes a simulated data synthesis model. In this case, step 203 specifically includes: receiving one or more target waveforms selected by the user from the simulated data pattern from the user operation platform; the target waveforms include at least: sine waves, square waves, triangle waves, and sawtooth waves. Then, the simulated data synthesis model is used to superimpose the one or more target waveforms to obtain a target simulated waveform, and sampling is performed from the target simulated waveform to obtain first synthetic data.

[0081] Based on the requirements for the statistical and business characteristics of the synthesized data, by selecting the appropriate mode and setting parameters such as wave amplitude and period, the simulated data synthesis model is used to superimpose the one or more target waveforms to obtain the target simulated waveform.

[0082] In summary, the method for generating synthetic data proposed in the embodiments of this specification uses different data synthesis models to generate synthetic data according to different types of synthetic materials, ensuring the consistency, reliability, and diversity of synthetic data with real-world data.

[0083] According to another embodiment, an apparatus for generating synthetic data is also provided. Figure 6A schematic block diagram of an apparatus for generating synthetic data according to one embodiment is shown. This apparatus can be deployed in any device, platform, or cluster of devices with computing and processing capabilities. Figure 6 As shown, the device 600 includes:

[0084] The acquisition unit 602 is configured to acquire at least one target synthetic material used to generate synthetic data. The types of target synthetic materials include real data collected by sensors of equipment during semiconductor manufacturing, data rules and / or business rules related to semiconductor manufacturing, and preset simulation data modes.

[0085] The first synthesis unit 603 is configured to process the target synthesis material using a corresponding data synthesis model based on the type of the target synthesis material to obtain corresponding first synthesis data. The amount of data in the first synthesis data is determined according to a preset first threshold. The data synthesis model includes an artificial intelligence data synthesis model corresponding to the real data, a rule data synthesis model corresponding to the data rules and / or business rules, and a simulated data synthesis model corresponding to the simulated data mode.

[0086] In some possible implementations, the device 600 further includes:

[0087] The weight determination unit 601 is configured to determine the weight values ​​of each quality assessment indicator based on the specific purpose of the first synthetic data; the quality assessment indicators include: business logic, data coverage, data similarity, correlation between data, and machine learning effectiveness.

[0088] The quality assessment unit 604 is configured to score the first synthetic data based on each quality assessment index, and to sum the scores according to their corresponding weight values ​​to obtain a quality assessment score.

[0089] The second synthesis unit 605 is configured to, when the quality assessment score is greater than a preset score threshold, process the target synthesis material using a corresponding data synthesis model based on the type of the target synthesis material to obtain corresponding second synthesis data, wherein the data volume of the second synthesis data is determined according to a preset second threshold; and the data volume of the second synthesis data is greater than the data volume of the first synthesis data.

[0090] According to another embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed in a computer, causes the computer to perform the methods described in any of the above embodiments.

[0091] According to another embodiment, a computing device is also provided, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, it implements the method described in any of the above embodiments.

[0092] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0093] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0094] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0095] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0096] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating synthetic data in a semiconductor manufacturing process, comprising: Acquire at least one target synthetic material used to generate synthetic data. The types of target synthetic materials include real data collected by sensors of equipment during semiconductor manufacturing, data rules and / or business rules related to semiconductor manufacturing, and preset simulation data patterns. Based on the type of the target synthetic material, the target synthetic material is processed using the corresponding data synthesis model to obtain the corresponding first synthetic data. The amount of data in the first synthetic data is determined according to a preset first threshold. The data synthesis model includes an artificial intelligence data synthesis model corresponding to the real data, a rule-based data synthesis model corresponding to the data rules and / or business rules, and a simulated data synthesis model corresponding to the simulated data pattern; wherein, when the target synthetic material includes real data, the data synthesis model includes an artificial intelligence data synthesis model; based on the type of the target synthetic material, the corresponding data synthesis model is used to process the target synthetic material to obtain the corresponding first synthetic data, including: Data analysis is performed on the real data to obtain its data characteristics, which include at least one of the following: statistical characteristics, data type of each field, data structure of each field, and logical analysis characteristics. Send the data features to the user operation platform; The system receives user-configured data rule constraints and / or business logic constraints for the synthesized data from the user operation platform. The data rule constraints include at least one of the following: data type, data distribution, data range, and primary / foreign key relationship. The business logic constraints include at least one of the following: the subordinate relationship between multiple fields in semiconductor manufacturing, and the constraint of a specific semiconductor physical mechanism model that one or more fields are subject to. At least the real data and the data rule constraints and / or business logic constraints are used to train each candidate model in the artificial intelligence data synthesis model. Based on the performance of each trained candidate model, one or more candidate models are determined as the first target model. The first target model is optimized using a preset hyperparameter optimization algorithm to obtain the second target model; The second target model is used to generate the first synthetic data.

2. The method of claim 1, wherein, The method further includes: Based on the specific purpose of the first synthetic data, the weight values ​​of each quality assessment indicator are determined; the quality assessment indicators include: business logic, data coverage, data similarity, correlation between data, and machine learning effectiveness. The first synthetic data is scored based on each quality assessment index, and the scores are weighted and summed according to their corresponding weight values ​​to obtain the quality assessment score. When the quality assessment score is greater than a preset score threshold, the target synthetic material is processed using a corresponding data synthesis model based on the type of the target synthetic material to obtain corresponding second synthetic data. The data volume of the second synthetic data is determined according to a preset second threshold; the data volume of the second synthetic data is greater than the data volume of the first synthetic data.

3. The method of claim 1, wherein, Before training the candidate models in the artificial intelligence data synthesis model using the real data, the method further includes: The real data is cleaned to remove or repair any unreasonable data; the unreasonable data includes at least one of the following: missing values, outliers, and duplicate values.

4. The method of claim 1, wherein, The target composite material includes data rules and / or business rules, and the data synthesis model includes a rule-based data synthesis model; based on the type of the target composite material, the corresponding data synthesis model is used to process the target composite material to obtain the corresponding first composite data, including: Receive user-configured data distribution type, shape factor, and size factor from the user operation platform; Based on the data rules and / or business rules, as well as the data distribution type, shape factor, and size factor, the rule-based data synthesis model is used to sample from the data distribution type, and perturbation noise data is added to the sampling results to obtain the first synthesized data.

5. The method of claim 1, wherein, The target composite material includes a simulated data pattern, and the data compositing model includes a simulated data compositing model; based on the type of the target composite material, the target composite material is processed using the corresponding data compositing model to obtain the corresponding first composite data, including: The system receives one or more target waveforms selected by the user from the simulation data mode from the user operation platform; the target waveforms include at least: sine wave, square wave, triangle wave, and sawtooth wave. The simulated data synthesis model is used to superimpose the one or more target waveforms to obtain a target simulated waveform, and sampling is performed from the target simulated waveform to obtain the first synthesized data.

6. An apparatus for generating composite data in a semiconductor manufacturing process, comprising: The acquisition unit is configured to acquire at least one target synthetic material used to generate synthetic data. The types of target synthetic materials include real data collected by sensors of equipment during semiconductor manufacturing, data rules and / or business rules related to semiconductor manufacturing, and preset simulation data modes. The first synthesis unit is configured to process the target synthetic material using a corresponding data synthesis model based on the type of the target synthetic material to obtain corresponding first synthetic data. The amount of data in the first synthetic data is determined according to a preset first threshold. The data synthesis model includes an artificial intelligence data synthesis model corresponding to the real data, a rule data synthesis model corresponding to the data rules and / or business rules, and a simulated data synthesis model corresponding to the simulated data mode. Wherein, when the target synthetic material includes real data, the data synthesis model includes an artificial intelligence data synthesis model; the first synthesis unit is further configured as follows: Data analysis is performed on the real data to obtain its data characteristics, which include at least one of the following: statistical characteristics, data type of each field, data structure of each field, and logical analysis characteristics. Send the data features to the user operation platform; The system receives user-configured data rule constraints and / or business logic constraints for the synthesized data from the user operation platform. The data rule constraints include at least one of the following: data type, data distribution, data range, and primary / foreign key relationship. The business logic constraints include at least one of the following: the subordinate relationship between multiple fields in semiconductor manufacturing, and the constraint of a specific semiconductor physical mechanism model that one or more fields are subject to. At least the real data and the data rule constraints and / or business logic constraints are used to train each candidate model in the artificial intelligence data synthesis model. Based on the performance of each trained candidate model, one or more candidate models are determined as the first target model. The first target model is optimized using a preset hyperparameter optimization algorithm to obtain the second target model; The second target model is used to generate the first synthetic data.

7. The apparatus of claim 6, wherein, The device further includes: The weight determination unit is configured to determine the weight values ​​of each quality assessment indicator based on the specific purpose of the first synthetic data; the quality assessment indicators include: business logic, data coverage, data similarity, correlation between data, and machine learning effectiveness. The quality assessment framework is configured to score the first synthetic data based on each quality assessment indicator, and then sum the scores according to their corresponding weight values ​​to obtain a quality assessment score. The second synthesis unit is configured to, when the quality assessment score is greater than a preset score threshold, process the target synthesis material using a corresponding data synthesis model based on the type of the target synthesis material to obtain corresponding second synthesis data, wherein the data volume of the second synthesis data is determined according to a preset second threshold; the data volume of the second synthesis data is greater than the data volume of the first synthesis data.

Images