A DRC rule making method, device, equipment and storage medium

Abstract

The application provides a DRC rule making method, device, equipment and storage medium, obtains a DR description file and an old version DRC configuration file; the obtained DR description file and the old version DRC configuration file are preprocessed to obtain basic data, a DRC field knowledge base is established based on the basic data; a DRC making model is established by using the DRC field knowledge base data, the parameters of the DRC making model are supervised and fine-tuned, the retrieval capability of the DRC making model on the DRC field knowledge base is enhanced, the output result of the DRC making model is optimized through Few-shot Learning, and the old version DRC configuration file is updated based on different DR data to output a corresponding new version DRC configuration file. The implementation of the application can improve the development quality of the DRC configuration file, ensure the accuracy of DRC generation in the chip design verification link, and shorten the DRC development time.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of chip research and design collaboration, and more particularly to a DRC rule making method, device, equipment and storage medium. BACKGROUND

[0002] In the field of integrated circuit design, as technology nodes continue to shrink, DTCO (Design Technology Co-Optimization) has become a key strategy to ensure efficient cooperation between design and process. The DTCO method allows circuit designers and process engineers to work closely at the early stages of design to optimize the PPAC (Power, Performance, Area, Cost) of integrated circuits.

[0003] In the traditional design process, DRC (Design Rule Check) is an important step to ensure that the design meets the requirements of a specific manufacturing process. However, with the development of technology, the design of integrated circuits has become more and more complex, especially on advanced technology nodes, design rules have become more stringent and numerous. In this case, later design modifications become more difficult and expensive. Therefore, how to quickly and accurately apply design rules at the early stages of the design process to prevent rule violations has become an urgent problem that the industry needs to solve.

[0004] The existing technology has encountered a bottleneck in dealing with this challenge. Although some tools and methods can provide some degree of DRC support at the early stages of design, they often cannot meet the needs of rapid iteration and efficient optimization. The performance of these tools is insufficient when dealing with complex designs, especially on advanced nodes, and their efficiency and accuracy problems are more obvious. In addition, as technology nodes continue to advance, the complexity and number of design rules are also increasing, making it difficult for existing DRC tools to keep up with the rapidly developing needs. SUMMARY

[0005] To solve the above technical problems, the present application provides a new methodology that overcomes the shortcomings of the prior art by using a more efficient design rule making and checking mechanism, improves the efficiency of the design process, shortens the product launch time, and reduces costs.

[0006] The core of the present application is to integrate DRC into the early stage of design, so that potential design and process conflicts can be found and solved earlier in the design process; by using advanced algorithms and automated script tools, the complexity of the rules is simplified, the number of rules that designers need to handle is reduced, while maintaining the quality of the design and the compatibility of the process. When the design does not meet the rules, automatic design adjustment can be performed to reduce manual intervention, improve efficiency, provide a real-time feedback system, allow PDK designers to see the modified results in real time, speed up the design iteration, promote collaboration between circuit designers and process engineers, and ensure that both parties can keep abreast of each other's needs and limitations to jointly promote design optimization.

[0007] The method optimizes the DTCO process by improving the DRC dedicated large language model, realizes the rapid generation and iteration of design rules to meet the design requirements of advanced technology nodes. This method is particularly suitable for complex IC design, which can ensure the design quality while speeding up the DRC formulation and optimization process, improving design efficiency, reducing rule violation risk and design cost.

[0008] The present application is realized by the following technical solutions.

[0009] The present application provides a DRC rule formulation method, comprising:

[0010] Obtain a DR specification file and an old version of a DRC configuration file, the DR specification file including DR rules;

[0011] Data preprocessing, including,

[0012] Preprocess the obtained DR specification file and the old version of the DRC configuration file using an automated script to obtain basic data,

[0013] Establish a DRC domain knowledge base based on the basic data;

[0014] Establish a DRC formulation model using the DRC domain knowledge base data, supervise and fine-tune the parameters of the DRC formulation model, and obtain a supervised and fine-tuned DRC formulation model,

[0015] Combine VSM technology and RAG technology to enhance the retrieval ability of the supervised and fine-tuned DRC formulation model on the DRC domain knowledge base,

[0016] Optimize the output results of the supervised and fine-tuned DRC formulation model through Few-shot Learning;

[0017] Update the old version of the DRC configuration file based on different DR data to output a corresponding new version of the DRC configuration file.

[0018] Further, the data preprocessing comprises:

[0019] Data collection, collected from the DR specification file and the old DRC configuration file through an automated script to collect data;

[0020] Data analysis, analyzing the collected data to determine the logical relationship and dependency of the data, and outputting the analysis result;

[0021] Rule connection, connecting the analysis result with a single specific DR rule;

[0022] Data matching, matching the DR rule in the DR specification file with the old DRC configuration file through pattern recognition technology.

[0023] Further, the matching of the DR rule in the DR specification file with the old DRC configuration file through pattern recognition technology is realized by the following formula:

[0024]

[0025] Wherein, n is the total number of DR rules, represents the i-th DR rule; is a rule set in the old DRC configuration file, and S is the total number of rules in the rule set in the old DRC configuration file, is the j-th rule in the old DRC configuration file; is a utility function for measuring the mapping quality of mapping , mapping , for each , there is , satisfying ; the matching function , wherein if matches , then , otherwise 0.

[0026] Further, the supervised fine-tuning is performed in the following manner:

[0027]

[0028] Wherein, is the number of labeled samples, is a loss function for measuring the difference between the model prediction result and the actual label , represents the feature vector of the p-th labeled sample, represents the p-th actual label, is a regularization parameter to prevent overfitting; is the dimension of the feature. Each feature vector is associated with its corresponding DRC rule label , where indicates whether the sample meets the PDK design rule, a classification model is trained using a supervised learning algorithm , which aims to minimize the prediction error , where represents the parameters of the model.

[0029] Further, the retrieval ability of the DRC formulation model on the DRC domain knowledge base is enhanced by the following way:

[0030]

[0031]

[0032] where, is a document in the DRC domain knowledge base, is the input query document; the retrieved document set is , is the probability of responding under the given query and the retrieved document set ;

[0033] a parameterized model is trained with parameters , the model learns after receiving a limited set of labeled sample combinations , E represents the total number of sample combinations in the labeled sample combination set , represents the second sample combination in the labeled sample combination set , can adapt to new PDK design rules by adjusting the parameters , this process is guided by the loss function , which measures the difference between the model output and the true label , the goal of model training is to minimize this loss function:

[0034]

[0035] Few-shot Learning and supervised fine-tuning work together to further improve the learning efficiency and accuracy of the model for new and specific design rules, under this condition, the model parameters are divided into two parts: basic parameters and adjustment parameters where In the large amount of DRC data, is adjusted by Few-shot Learning, the parameter update of the model is expressed as:

[0036]

[0037] The retrieval process is regarded as a query function , the output of the model also includes the relevant information retrieved from the DRC domain knowledge base:

[0038]

[0039]

[0040] where K represents the DRC domain knowledge base, The function represents the process generated by the model, The function represents the process based on the input Retrieving information from the DRC domain knowledge base K.

[0041] Further, the different DR data output corresponding new version DRC configuration file based on the specific matter process, including the following steps:

[0042] Suppose the model has a new input of DRC configuration file rule set from the new PDK , the ith DRC configuration file rule contains a corresponding design parameter , the design parameter , for each ith , extract its corresponding feature vector , through the feature extraction function , that is:

[0043] Apply clustering algorithm to analyze the feature vector set to identify common patterns: where represents the identified design rule pattern set;

[0044] For new PDK requirements, expressed as a set of design parameters , define an adaptation function ɑ to adjust the existing patterns to meet the new requirements:

[0045] Finally, define an optimization O process to refine the rules to ensure its effectiveness and optimality under the new PDK: .

[0046] The application further provides a DRC rule making device, comprising

[0047] a data preprocessing module, configured to obtain basic data obtained after preprocessing of the DR specification file and the DRC configuration file;

[0048] a DRC domain knowledge base establishing module, configured to create the basic data obtained after preprocessing of the DR specification file and the DRC configuration file as a DRC domain knowledge base;

[0049] a DRC making model constructing module, configured to train and optimize a DRC making model;

[0050] a new version DRC configuration file outputting module based on different DR data sources, configured to output a corresponding new version DRC configuration file according to input DR data of a specific process node.

[0051] The application further provides an electronic device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the DRC rule making method of any one of the above when executing the program.

[0052] The application further provides a non-transitory computer readable storage medium, which stores a computer program, wherein the computer program is executed by a processor to implement the DRC rule making method of any one of the above.

[0053] The application has the following beneficial effects:

[0054] (1) The method optimizes the DTCO process by improving the DRC special large language model, realizes the rapid generation and iteration of the design rules, and meets the design requirements of advanced technology nodes. The method is particularly suitable for complex IC design (Integrated Circuit Design), can accelerate the DRC making and optimization process while ensuring the design quality, improve the design efficiency, reduce the risk of rule violation and design cost.

[0055] (2) Improved efficiency in DRC formulation and optimization. By introducing automated tools and advanced algorithms, the efficiency of the DRC formulation and optimization process is significantly improved. Traditional DRC methods rely on manual adjustments and iterative processes, which are time-consuming and prone to errors. The present invention utilizes automated algorithms to quickly and accurately identify potential design issues and provide optimization suggestions, reducing the need for human intervention. This efficiency improvement allows PDK designers to respond more quickly to design rule changes, speeding up the design verification process and reducing the time from concept to market. At the same time, the automated DRC optimization process also reduces the workload of PDK designers, allowing them to focus more on innovation and other key aspects of design, thereby improving overall design productivity.

[0056] (3) Enhanced adaptability and migration of DRC configuration code generation for evolving technology nodes. The modular design rule formulation method provides greater scalability and flexibility for different types and sizes of design projects. In traditional DRC processes, DRC configuration code is often fixed for specific chip processes, which limits the applicability of rules between different projects. The present invention allows design rules to exist in a modular form, which can be quickly adjusted and reorganized according to the specific needs of the project. This flexibility not only applies to diverse design projects, but also adapts to rapidly changing manufacturing processes and technology standards.

[0057] (4) Reduced potential risk of DRC violations through predictive analysis. By introducing advanced large language prediction models, potential DRC conflict issues can be predicted in the early stages of the design process. This predictive analysis allows PDK designers to make necessary modifications before submitting the final design, significantly reducing the likelihood of violations. This not only reduces the number of rework, but also shortens the overall design cycle, accelerating the time to market. In addition, this method helps to identify and solve potential problems in the subsequent manufacturing process in advance, thereby reducing the overall project risk and ensuring the reliability and performance of the final product. Through such predictive analysis, the PDK design team can more accurately plan the design process, optimize resource allocation, and improve the efficiency of the entire design and manufacturing process. By improving the scalability and flexibility of DRC code generation, the present invention supports more extensive design innovation and promotes further improvement in PDK design efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 is a general flowchart of the present invention.

[0059] Figure 2 is a block diagram of the design rule automatic generation module in the present invention.

[0060] Figure 3 Flow chart for formulating model of DRC in embodiments of the present application.

[0061] Figure 4 For input 40nm process node DR file content, automatically generate DRC configuration code content.

[0062] Figure 5 Schematic diagram of man-machine interface of the present application.

[0063] Figure 6 Rule verification and feedback mechanism diagram in the present application.

[0064] Figure 7 Structure schematic diagram of DRC rule formulation device provided by the present application;

[0065] Figure 8 Structure schematic diagram of electronic device provided by the present application. DETAILED DESCRIPTION

[0066] The technical solutions of the present application are further described below, but the scope of protection is not limited to the description.

[0067] A DRC rule formulation method comprises the following steps:

[0068] Obtain a DR specification file and an old version DRC configuration file, the DR specification file comprising DR rules; data preprocessing, comprising, preprocessing the obtained DR specification file and the old version DRC configuration file by an automatic script to obtain basic data, and establishing a DRC field knowledge base based on the basic data; establishing a DRC formulation model by using the DRC field knowledge base data, supervising and fine-tuning parameters of the DRC formulation model, combining VSM technology and RAG technology to enhance the retrieval ability of the supervised and fine-tuned DRC formulation model on the DRC field knowledge base, optimizing the output result of the supervised and fine-tuned DRC formulation model by Few-shot Learning; updating the old version DRC configuration file based on different DR data to output a corresponding new version DRC configuration file.

[0069] By the new version DRC configuration file, the same operation above can be performed to obtain a new DRC field knowledge base, and further obtain a new DRC formulation model.

[0070] The data preprocessing includes: data collection, collecting data from the DR specification file and the old DRC configuration file through an automated script; data analysis, analyzing the collected data to determine the logical relationship and dependency of the data, and outputting the analysis result; rule connection, connecting the analysis result with a single specific DR rule; data matching, matching the DR rule in the DR specification file with the old DRC configuration file through pattern recognition technology.

[0071] The present application mainly automatically parses the rules in the DR specification file through an automated script (such as a Python or Shell script), and intelligently matches these rules with existing DRC configuration files. This process significantly reduces manual intervention, improves the efficiency and accuracy of rule making. The established DRC domain knowledge base integrates scattered design rule and configuration file information, providing high-quality data support for subsequent large language model training and retrieval.

[0072] , n is the total number of DR rules, represents the i-th DR rule; is a rule set in the old DRC configuration file, and S is the total number of rules in the rule set in the old DRC configuration file, is the j-th rule in the old DRC configuration file; is a utility function for measuring the mapping quality of is the number of successfully matched rules or the accuracy of matching.

[0073] The goal of the automated script is to find a mapping For each , there is that satisfies ; the matching function where if matches , then , otherwise 0.

[0074] Therefore, the task of the automated script can be formulated as the following optimization problem:

[0075]

[0076] Using this formula algorithm, the data preprocessing tries to maximize the matching utility while ensuring that each rule can find at least one matching configuration .

[0077] The parameters of the large language model are adjusted using SFT (Supervised Fine-Tuning) technology, and the VSM (Vector Space Model) and RAG (Retrieval-Augmented Generation) technologies are used to enhance the retrieval mapping ability of the DRC domain knowledge base data. The Few-shot Learning (few-shot learning) optimizes the content of the DRC configuration file output by the model.

[0078] Through SFT technology, the parameters of the large language model are adjusted according to the characteristics of the DRC configuration file, making it more suitable for specific application scenarios. Combined with VSM and RAG technologies, the model's retrieval and mapping capabilities for DRC domain knowledge bases are enhanced, improving the relevance and accuracy of the generated results. Meanwhile, through Few-shot Learning, the model quickly adjusts its parameters after receiving limited labeled DRC rule samples, enabling it to accurately adapt to and reflect the design rule requirements of a specific PDK even in the face of insufficient data. The comprehensive application of these technologies enables the model to automatically generate high-accuracy DRC configuration files based on input DR rules, breaking through the bottlenecks of traditional methods.

[0079] Extract the feature vector of each labeled sample where is the sample index, and is the dimension of the feature. Each feature vector is associated with its corresponding DRC rule label , where indicates whether the sample meets the PDK design rules. Using a supervised learning algorithm, a classification model is trained to minimize the prediction error , where represents the model's parameters. The objective function is defined by the following formula:

[0080]

[0081] where is the number of labeled samples, is the loss function, which measures the difference between the model's prediction and the actual label , is a regularization parameter to prevent overfitting.

[0082] Through an iterative optimization algorithm, the model parameters , until convergence to a minimum error or reaching a predetermined number of iterations. Through the above steps, the SFT technique can fine-tune the parameters according to the design rule requirements of a specific PDK, thereby improving the accuracy and efficiency of the automatic generation method in DRC rule detection.

[0083] VSM and RAG techniques combine the generation and retrieval capabilities of large language models to improve the efficiency and accuracy of querying DRC domain knowledge base data. The VSM process first converts queries and documents into vectors. These vectors are represented in a high-dimensional space, where each dimension corresponds to a specific term. For a document and a query , their vector representations are and respectively. The similarity between a document and a query is usually calculated by the cosine similarity of their vectors:

[0084]

[0085] where the dot product represents the inner product of two vectors, and are the Euclidean norms of the vectors.

[0086] The RAG technique combines vector space model-based retrieval and transformer-based language model generation. For a given input query , RAG first retrieves relevant documents in the DRC domain knowledge base using VSM. This is done by calculating the similarity between the query vector and the vector of all documents in the knowledge base. The retrieved document set is .

[0087] RAG then inputs the retrieved documents as context information into the SFTed DRC-specific model. This model combines the retrieved context and the original query to generate a response . The response generated by RAG can be represented as:

[0088]

[0089] where is the probability of the response given the query and the retrieved document set .

[0090] Finally, RAG adopts a fusion strategy to combine the retrieval and generation results to produce the final output.

[0091] Through this combination of retrieval and generation, RAG can leverage the large amount of information in the DRC domain knowledge base to improve the relevance and accuracy of the answers while maintaining the fluency and coherence of the language model-generated responses.

[0092] With Few-shot Learning, the model can quickly adjust its parameters after receiving a limited set of labeled DRC rule samples, allowing it to accurately adapt to and reflect the design rule requirements of a specific PDK without the need for large amounts of training data.

[0093] Assuming the training of a parameterized model with parameters , the model learns after receiving a limited set of labeled sample combinations , where E represents the total number of sample combinations in the labeled sample combination set , and represents the second sample combination in the labeled sample combination set In the Few-shot Learning scenario, the model can adapt to new PDK design rules by adjusting the parameters , which can be guided by the loss function , which measures the difference between the model output and the true label . The goal of model training is to minimize this loss function:

[0094]

[0095] Few-shot Learning works in conjunction with supervised fine-tuning to further improve the learning efficiency and accuracy of the model for new and specific design rules. Under this condition, the model parameters are divided into two parts: the base parameters and the adjustment parameters , where is trained on a large amount of DRC data, is adjusted through Few-shot Learning, and therefore, the parameter update of the model can be expressed as:

[0096]

[0097] Combining VSM and RAG techniques, Few-shot Learning enables the model to query the DRC domain knowledge base even for rarely seen rule information, improving the efficiency and accuracy of the query through effective retrieval and generation operations. The retrieval process can be seen as a query function , the output of the model also includes the retrieved relevant information:

[0098]

[0099]

[0100] where K represents the DRC domain knowledge base, the function represents the process of model generation, the function represents the process of input retrieving information from the DRC domain knowledge base K.

[0101] The method realizes strong migration of DRC rapid formulation based on different DR data sources.

[0102] After the model training and optimization are completed, the proposed method has good migration and can adapt to DR data from different sources, realize the rapid formulation of DRC configuration files, and meet the development needs of different PDKs. This means that the method is not limited to specific DR data sources or PDKs and can adapt to different design rules and process requirements. By analyzing and comparing DR data from different sources, the method can identify common design rule patterns and quickly adapt to new PDK requirements, thereby realizing the rapid formulation and migration of DRC rules in different design processes. This feature improves the applicability and practical value of the method.

[0103] Assume that the model has newly input a set of DRC configuration file rules from a new PDK , the i-th DRC configuration file rule contains a corresponding design parameter , the design parameter , for each i-th , extract its corresponding feature vector , through the feature extraction function , that is:

[0104] Apply clustering algorithm to analyze the feature vector set to identify common patterns: where represents the set of identified design rule patterns;

[0105] For new PDK requirements expressed as a set of design parameters , define an adaptation function a to adjust the existing patterns to meet the new requirements:

[0106] Finally, define an optimization O process to refine the rules , ensure its effectiveness and optimality under the new PDK: .

[0107] The basic step is to collect, analyze, rule connect and match the DR specification file with the existing DRC configuration file through an automated script, and establish a knowledge base in the DRC field. This knowledge base provides necessary data support for subsequent large language model training and reasoning.

[0108] Using the established knowledge base data, the large language model is supervised fine-tuned (SFT) to adapt to the generation task of DRC configuration file. Combined with VSM and RAG technology, the retrieval ability of the model to the knowledge base is enhanced, and the model output is optimized through Few-shot Learning.

[0109] The application stage of the method, after the model training and optimization are completed, uses the trained model to quickly formulate the DRC configuration file in actual application. High migration means that the method can adapt to different DR data sources, and realize fast DRC formulation across process nodes and different design rules.

[0110] Through the combination and application of the above method, the automatic generation of DRC configuration file method proposed in the application not only improves the generation efficiency of the configuration file, but also ensures the high accuracy of the generated file, thereby solving the bottleneck problem of DRC configuration file development in the chip design verification link.

[0111] Taking the automatic generation of DR and the corresponding DRC configuration content of the 40nm process node as an example, a DRC rule formulation method of the application is used to realize the flow. As shown in Figure 1 , the overall overview of the entire DRC fast formulation process is shown, including data preprocessing, knowledge base construction, model training, human-computer interaction and rule verification.

[0112] The process starts from "start", first data preprocessing, this step includes automatic data collection, data analysis, rule connection and rule matching. Then judge whether the data meets the knowledge base requirements, if not, the process is ended, if yes, create PDK development kit DRC field knowledge base. Then build a DRC (design rule check) formulation model based on a large language model, this process includes SFT training, VSM + RAG data retrieval mapping and Few-shot Learning output optimization. Then use different DR data sources to formulate DRC quickly with strong migration, and check whether the DRC formulation model representation is good, if the model representation is not good, the process is ended, if the representation is good, enter the last step "end". The whole process is successfully ended when the model representation is good.

[0113] As shown in Figure 2As shown, data preprocessing for the DRC large language knowledge base includes data collection, analysis, rule connection and matching submodules. Data is collected from existing design databases and old DRC configuration files through automated scripts. The collected data is analyzed to determine the logical relationship and dependency of design rules. The analysis results are connected with specific DR rules to provide basic data for knowledge base construction. DR rules are matched with existing DRC configuration files through pattern recognition technology.

[0114] As shown in Figure 3 , the training and optimization process of the DRC formulation model is described in detail. The SFT technology is used to adjust part of the parameters of the large language model to adapt to the characteristics of the DRC configuration file. VSM and RAG technologies are used to enhance the data retrieval capability. Few-shot Learning scheme is used to train the model based on a small number of samples to optimize the output configuration code content.

[0115] As shown in Figure 4 , the DR specific content of the 40nm process node is input into the trained DRC formulation model in text form. The model automatically outputs the corresponding DRC configuration code according to the input DR rules.

[0116] As shown in Figure 5 , the process diagram of human-computer interaction is shown. Manually interact with the automated tool through the graphical interface to formulate and optimize the design rules. Specific design parameters or rule requirements are input through the graphical interface. The system provides real-time feedback, and the PDK designer adjusts the input parameters and rules according to the feedback.

[0117] As shown in Figure 6 , the process of verifying and adjusting design rules according to feedback is shown. The system applies the generated DRC configuration file to the actual design for verification. According to the verification result, the configuration file is adjusted and optimized to ensure its accuracy and effectiveness. A feedback link is set up to collect the suggestions of PDK designers and the verification results of the system, and the model is continuously optimized.

[0118] The DRC rule making method provided by the application can automatically generate a corresponding DRC configuration file according to the input DR rule content, thereby reducing the error rate and workload of manual configuration file writing.

[0119] The DRC rule making device provided by the application is described as follows, Figure 7 The DRC rule making device described below can be referred to the DRC rule making method described above.

[0120] The DRC rule making device comprises

[0121] The data preprocessing module 710 is configured to obtain the basic data obtained after preprocessing the DR specification file and the DRC configuration file;

[0122] The DRC domain knowledge base module 720 is configured to create the basic data obtained after preprocessing the DR specification file and the DRC configuration file as a DRC domain knowledge base;

[0123] The DRC making model module 730 is configured to train and optimize a DRC making model;

[0124] The DRC making model module 730 is configured to train and optimize a DRC making model;

[0125] Figure 8 An example of an entity structure diagram of an electronic device is shown in Figure 8 The electronic device can comprise a processor 810, a communications interface 820, a memory 830 and a communications bus 840, wherein the processor 810, the communications interface 820 and the memory 830 can communicate with each other through the communications bus 840. The processor 810 can call the logical instructions in the memory 830 to execute the DRC rule making method, which comprises:

[0126] The method comprises the following steps: obtaining DR specification files and old version DRC configuration file data preprocessing, including, preprocessing the obtained DR specification files and the DRC configuration files by using an automatic script to obtain basic data; establishing a DRC field knowledge base, including, establishing the DRC field knowledge base by using the basic data obtained in the data preprocessing step; constructing a DRC formulation model, including, supervising and fine-tuning the parameters of the DRC formulation model by using the DRC field knowledge base data, combining VSM technology and RAG technology to enhance the retrieval ability of the DRC formulation model to the DRC field knowledge base, and optimizing the output of the DRC formulation model through Few-shot Learning; outputting corresponding new version DRC configuration files based on different DR data sources, including, inputting DR data of a specific process node into the trained DRC formulation model, and the DRC formulation model automatically outputs corresponding new version DRC configuration files according to the input DR data.

[0127] The application further provides a non-transitory computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the DRC rule formulation method provided by each of the methods, and the method comprises the following steps:

[0128] The method comprises the following steps: obtaining DR specification files and old version DRC configuration file data preprocessing, including, preprocessing the obtained DR specification files and the DRC configuration files by using an automatic script to obtain basic data; establishing a DRC field knowledge base, including, establishing the DRC field knowledge base by using the basic data obtained in the data preprocessing step; constructing a DRC formulation model, including, supervising and fine-tuning the parameters of the DRC formulation model by using the DRC field knowledge base data, combining VSM technology and RAG technology to enhance the retrieval ability of the DRC formulation model to the DRC field knowledge base, and optimizing the output of the DRC formulation model through Few-shot Learning; outputting corresponding new version DRC configuration files based on different DR data sources, including, inputting DR data of a specific process node into the trained DRC formulation model, and the DRC formulation model automatically outputs corresponding new version DRC configuration files according to the input DR data.

[0129] The above only describes the preferred embodiments of the application, and any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. A method of DRC rule formulation, characterized by, The method comprises the following steps: acquiring a DR specification file and an old DRC configuration file, wherein the DR specification file comprises DR rules; data preprocessing, comprising: preprocessing the acquired DR specification file and the old DRC configuration file by using an automatic script to obtain basic data; establishing a DRC domain knowledge base based on the basic data; establishing a DRC formulation model by using the DRC domain knowledge base data, performing supervised fine-tuning on parameters of the DRC formulation model, and obtaining a supervised fine-tuned DRC formulation model; enhancing the search capability of the supervised fine-tuned DRC formulation model on the DRC domain knowledge base by combining VSM technology and RAG technology; optimizing the output result of the supervised fine-tuned DRC formulation model by Few-shot Learning; updating the old DRC configuration file based on different DR data to output a corresponding new DRC configuration file.

2. The method of claim 1, wherein, The data preprocessing comprises: data acquisition, collecting data by automatically collecting from the DR specification file and the old DRC configuration file through an automatic script; data analysis, analyzing the collected data to determine the logical relationship and dependency of the data, and outputting an analysis result; rule connection, connecting the analysis result with a single specific DR rule; data matching, matching the DR rules in the DR specification file with the old DRC configuration file by using pattern recognition technology.

3. The method of claim 2 wherein the DRC rule is determined by, The matching of the DR rules in the DR specification file with the old DRC configuration file by using pattern recognition technology is realized by the following formula: wherein, n is the total number of DR rules, represents the ith DR rule; is a rule set in the legacy DRC configuration file, S is the total number of rules in the rule set in the legacy DRC configuration file, is the jth rule in the legacy DRC configuration file; is a utility function for measuring the mapping quality of mapping is a mapping function, mapping , for each , there is , satisfying ; a matching function , wherein if matches , then , otherwise 0.

4. The method of claim 1, wherein, The supervised fine-tuning is performed in the following manner: wherein, is the number of labeled samples, is a loss function that measures the difference between the model's prediction and the actual label, represents the feature vector of the pth labeled sample, represents the pth actual label, is a regularization parameter to prevent overfitting; and is the dimension of the feature, each feature vector is associated with its corresponding DRC rule label wherein represents whether the sample meets the PDK's design rules, a classification model is trained using a supervised learning algorithm , the model aims to minimize the prediction error wherein represents the model's parameters.

5. The method of claim 1 wherein, The search capability of the DRC formulation model on the DRC domain knowledge base is enhanced in the following manner: wherein, is a document in the DRC domain knowledge base, is an input query document; the retrieved document set is , is the probability of responding to a given query under the condition of the retrieved document set ; Training a parameterized model With parameters The model receives a limited set of labeled sample combinations. Then learning is performed, where E represents the set of labeled sample combinations. The total number of sample combinations Represents the set of labeled sample combinations The second sample combination, in the Few-shot Learning scenario, the model Able to adjust parameters To adapt to the new PDK design rules, this process uses a loss function. The loss function provides guidance and measures the model output. With real labels The difference between them, the goal of model training is to minimize this loss function: Few-shot Learning works with supervised fine-tuning to further improve the learning efficiency and accuracy of the model for new and specifically designed rules. Under this condition, the model parameters are divided into two parts: basic parameters and adjustment parameters , wherein in a large amount of DRC data for training, is adjusted by Few-shot Learning. The parameter update of the model is expressed as: The search process is viewed as a query function The output of the model also includes the relevant information retrieved from the DRC domain knowledge base V. where K represents a DRC domain knowledge base, the process of function representation model generation, function representation is based on input the process of retrieving information from the DRC domain knowledge base K.

6. The method of claim 1 wherein, The specific process of outputting the corresponding new DRC configuration file based on different DR data comprises the following steps: Set model new inputed DRC configuration file rule set from new PDK , the ith DRC configuration file rule contains a corresponding design parameter , the design parameter , for each ith , extract its corresponding feature vector , through the feature extraction function , that is: Applying a clustering algorithm to the set of feature vectors to identify common patterns: wherein represents the identified set of design rule patterns; For a new PDK requirement, expressed as a set of design parameters Define an adaptation function a to adjust the existing pattern to meet the new requirement: Finally, an optimization O procedure is defined to refine the rules , ensuring their validity and optimality under the new PDK: .

7. A DRC rule making apparatus characterized by comprising: The method comprises a data preprocessing module for obtaining basic data after preprocessing the DR specification file and the old DRC configuration file; a DRC domain knowledge base establishment module for creating the basic data obtained after preprocessing the DR specification file and the DRC configuration file as a DRC domain knowledge base; a DRC formulation model construction module for training and optimizing a DRC formulation model: establishing a DRC formulation model by using the DRC domain knowledge base data, performing supervised fine-tuning on parameters of the DRC formulation model, and obtaining a supervised fine-tuned DRC formulation model, enhancing the search capability of the supervised fine-tuned DRC formulation model on the DRC domain knowledge base by combining VSM technology and RAG technology, optimizing the output result of the supervised fine-tuned DRC formulation model by Few-shot Learning, a new DRC configuration file output module based on different DR data sources for outputting a corresponding new DRC configuration file according to input DR data of a specific process node.

8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to realize the DRC rule formulation method according to any one of claims 1 to 6. 9.A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the DRC rule formulation method according to any one of claims 1 to 6.

Images