A layout design rule generation method, device and medium based on a multi-modal agent technology

Abstract

The application discloses a layout design rule generation method and device based on a multi-modal intelligent agent technology and a medium, relates to the field of electric digital data processing, and converts a design rule manual into a structured text format and divides the structured text format according to process levels. A visual model is used to perform semantic extraction on a schematic diagram in the manual, generate an image description text, and fuse the image description text with an original rule text. A candidate function set is screened from a function library, historical experience rule library is combined, and a function or function combination required by a current rule is decided. Detailed usage information is acquired, process level auxiliary constraints and global instructions are retrieved from the manual, code specifications are retrieved from the experience knowledge base, all knowledge is enhanced to guide rule semantic restatement, image information selection, function parameter analysis and pre-output self-checking, and layout design rule checking code is generated. The application solves the problems of low manual conversion efficiency and errors, and has the advantages of high generation efficiency, accurate results and good interpretability.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and more particularly to computer-aided design of integrated circuits / electronic design automation, specifically a layout design rule generation method, device, and medium based on multimodal intelligent agent technology. Background Technology

[0002] As integrated circuit manufacturing processes continue to evolve towards smaller nodes, higher demands are placed on Design Rule Checking (DRC). DRC is a core step in the physical verification of integrated circuits. Its purpose is to automatically check the geometry, interlayer relationships, dimensional spacing, and extension constraints in the layout based on the design rule manual provided by the process manufacturer. This identifies and marks non-compliant areas that do not meet manufacturing process requirements, ensuring the layout can be successfully fabricated. With increasing chip complexity and the exponential growth in the number of design rules in advanced processes, the design rule manual itself has become increasingly large and complex. A typical design rule manual usually contains hundreds or even thousands of rules, including not only constraints described in natural language but also numerous tabular parameters, geometric diagrams, annotations, and derivation layer reuse relationships, function combination and call relationships, and context dependencies between different rules. This information is presented in a mix of modalities such as text, images, and tables, posing a significant challenge to the engineering implementation of the rules.

[0003] In existing technologies, the development of DRC rules primarily relies on manual methods. Engineers first need to thoroughly read and understand the natural language descriptions and geometric diagrams in the Design Rule Manual, and then, based on their own physical verification experience and programming knowledge of specific electronic design automation (EDA) software, convert the constraint information in the manual line by line into rule checking code executable by the software. To alleviate the inefficiency of manual coding, some research has proposed auxiliary methods based on natural language processing (NLP) technology, such as extracting key information from the manual and filling it into predefined code templates. In recent years, with the development of large language models, some scholars have also attempted to use large models to automatically generate relatively simple design rules, for example, by prompting engineering guidance models to output rule code.

[0004] However, these existing solutions all have significant drawbacks. First, manual rule writing heavily relies on the accumulated experience of experts. Understanding and debugging a single rule can often take tens of minutes or even hours, resulting in a long development cycle and making it prone to omissions or misinterpretations due to misunderstandings or oversights. Second, different physical verification software have completely different function libraries and syntax rules, meaning manually written rule code is usually not reusable across platforms, further exacerbating repetitive work. Furthermore, template-based methods using natural language processing lack the ability to understand image information, while many crucial geometric constraints, measurement methods, directional relationships, and prohibited examples in the design rule manual are precisely conveyed through diagrams; textual descriptions alone cannot fully capture this information. Simultaneously, existing methods for generating rules using large models directly input complete rule or function manuals into the model, leading to an excessive burden on the model context. Moreover, general-purpose large models lack specialized knowledge of layout design domains and specific electronic design automation tools, resulting in code that often contains syntax errors or logical gaps, leading to poor executability.

[0005] Therefore, there is an urgent need to propose a technical solution that can convert rule text descriptions, image diagrams, and specialized knowledge into high-quality, executable design rule checking code. Summary of the Invention

[0006] This invention overcomes the shortcomings of the prior art and provides a method, device and medium for generating layout design rules based on multimodal intelligent agent technology.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: Firstly, the present invention provides a layout design rule generation method based on multimodal intelligent agent technology, comprising the following steps:

[0008] S1. Convert the source file of the design rule manual containing text, tables and images into a structured text format and divide it according to the process level;

[0009] S2. Use a visual model to extract semantics from the schematic diagrams in the segmented file to generate image description text; fuse the image description text with the original rule text, with the original rule text as the priority and the image description text as supplementary information, to obtain an enhanced rule description.

[0010] S3. Based on the enhanced rule description, a set of candidate functions is selected from the function library through vector similarity retrieval, and combined with the historical experience rule library, the first-level reasoning model is used to determine the function or combination of functions required to generate the current rule.

[0011] S4. Query the function usage database to obtain the usage information of the function or function combination determined by the decision, and retrieve the auxiliary constraints and global descriptions of the process level involved in the current rule from the design rule manual, and retrieve the relevant code specifications from the experience knowledge base. All the retrieved information is used as knowledge enhancement and input into the second-level reasoning model to guide it to reason according to the preset thinking chain, and finally generate executable layout design rule check code.

[0012] In a preferred embodiment of the present invention, step S2 includes the following sub-steps:

[0013] S21. For images in the same process layer, maintain a mapping table between legend colors and process layer names, and extract the rule names contained in the layer from the document table to establish a rule candidate set.

[0014] S22. Input the legend mapping, image file and rule name into the visual model, and extract the effective information in the diagram through question and answer. The effective information includes: involved layer, measurement object, measurement method, prohibition symbol and formula expression.

[0015] S23. When there is a semantic conflict between the rule text description and the diagram information, the rule text description shall be given first priority, the quantitative relationship in the diagram shall be given as a highly reliable supplement, and the directional relationship in the diagram shall be given as the second priority, and the reasoning model shall make the choice.

[0016] In a preferred embodiment of the present invention, step S3 includes the following sub-steps:

[0017] S31. On the one hand, based on TF-IDF vectors and cosine similarity, the similarity between the rule description text and the function names and function descriptions in the function library is calculated to obtain the first candidate function set;

[0018] On the other hand, the rule description text is matched with the anonymized rule description in the historical experience knowledge base, and the empirical rules with the highest similarity and their function combinations are extracted as the second candidate function set.

[0019] S32. Merge the first candidate function set and the second candidate function set, and pass the names of the functions, usage descriptions and relevant historical experience examples as input to the first-level inference model, which will then decide on the final required function or combination of functions.

[0020] In a preferred embodiment of the present invention, in step S4, the function usage database is constructed in the following manner: the function usage manual of the physical verification software is preprocessed, indexed by function name, and the function usage description, format specification and parameter description are used as values ​​to establish an SQLite database.

[0021] In a preferred embodiment of the present invention, in step S4, the knowledge enhancement further includes: extracting the process level involved by parsing the name of the current rule, retrieving auxiliary constraint information related to that level in the entire document, and retrieving global description information from the design rule manual, and inputting the above information as supplementary knowledge into the second-level reasoning model.

[0022] In a preferred embodiment of the present invention, in step S4, the construction and retrieval of the experience knowledge base includes: checking the code file according to existing layout design rules, and constructing it after anonymizing the specific process level; during retrieval, it is performed by calculating the vector similarity of the rule description text or by keyword matching.

[0023] In a preferred embodiment of the present invention, in step S4, the preset thought chain includes the following reasoning steps executed in sequence: rule semantic understanding and constraint target restatement, selection of image summary information and empirical knowledge, applicability analysis of candidate functions, understanding of supplementary concept information, analysis of the selection or rejection of function parameters one by one, judgment of the reuse of derived layers, and self-check of common problems before output.

[0024] In a preferred embodiment of the present invention, steps S3 and S4 adopt a two-level inference model architecture: the first-level inference model receives the simplified candidate function information and uses it to decide the required function or combination of functions; the second-level inference model receives the decision result of the first level and queries to obtain the detailed function usage information and various domain knowledge of the result for the final code generation.

[0025] In a second aspect, embodiments of the present invention provide an electronic device, comprising: at least one processor; and a memory communicatively connected to at least one of the processors;

[0026] The memory stores a computer program that is executed by at least one of the processors, such that the at least one processor is able to execute the layout design rule generation method based on multimodal intelligent agent technology as described above.

[0027] Thirdly, embodiments of the present invention provide a computer-readable storage medium storing computer instructions, which are used to cause a processor to execute the layout design rule generation method based on multimodal intelligent agent technology described above.

[0028] This invention addresses the shortcomings of the prior art and has the following beneficial effects:

[0029] (1) This invention provides a layout design rule generation method based on multimodal intelligent agent technology. By converting the schematic diagram in the design rule manual into a natural language description using a visual model and fusing it with the original rule text with text priority, it can overcome the shortcomings of the pure text method in obtaining visual information such as geometric measurement direction, prohibition symbols and interlayer contact relationship. In this way, the reasoning model can obtain text constraints and graphic semantics at the same time, and can fully understand the measurement object, measurement method and exception of the rule. Compared with the existing methods that only rely on natural language processing or directly input large models, this invention can effectively avoid the deviation in rule semantic understanding caused by the lack of image information, thereby improving the logical integrity of the generated rule code.

[0030] (2) The present invention adopts a two-level inference model architecture. The first-level model first filters out a set of candidate functions from hundreds of functions based on vector similarity retrieval, and then decides the function or combination of functions required for the current rule. The second-level model only obtains the detailed function usage information corresponding to the decision result to generate code, thereby greatly reducing the length of input information required for each inference, so that the model can focus on the functions and grammatical details most relevant to the current rule. Compared with the practice of directly inputting a complete manual into the model in the prior art, the present invention not only reduces the processing burden of the model, but also reduces the risk of logical confusion or illusory output due to excessively long context, thereby ensuring the grammatical correctness and executableness of the generated code.

[0031] (3) By constructing a function usage database, an engineering experience knowledge base, and a process-level auxiliary constraint retrieval mechanism, this invention injects the dedicated function format, historical best practice rules, and global process descriptions of the physical verification software as knowledge enhancement content into the inference model. This can supplement the model's shortcomings in dedicated knowledge, enabling it to call functions in the correct format and follow verified code specifications. Compared with the high cost of training dedicated large models, this invention achieves similar generation quality with a low-cost knowledge retrieval injection method, thereby improving the usability and reliability of rule code in the actual physical verification process.

[0032] (4) The present invention divides the design rule manual into process layers and performs image extraction, function decision and code generation independently for each rule. This enables the model to focus the context of each rule on the relevant process layer and constraint information. Compared with the solution of processing the entire document at once, the present invention effectively avoids the attention shift problem and supports parallel processing of rules at different levels, thereby improving the efficiency and scalability of the overall rule generation. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a schematic diagram of the agent interaction in the generation method of the present invention;

[0035] Figure 2 This is a schematic diagram of rules GCON.9 and GCON.14 of this invention;

[0036] Figure 3 This is a diagram showing the prompt words of the reasoning model when generating rules in this invention;

[0037] Figure 4 A schematic diagram of an electronic device structure that can be used to implement Embodiment 1 of the present invention is shown.

[0038] In the diagram: 10. Electronic device; 11. Processor; 12. Read-only memory; 13. Random access memory; 14. Bus; 15. I / O interface; 16. Input unit; 17. Output unit; 18. Storage unit; 19. Communication unit. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0041] Example 1

[0042] This embodiment uses the open-source inference model DeepSeek-V3.1 and the open-source vision model Qwen-VL-Instruct as examples, and the FreePDK3 design rule manual as the test object. This manual contains multiple process layers, such as NW active region, ACT active region, BPR barrier layer, GCON contact hole layer, GCUT cutting layer, DUMMY virtual layer, etc., totaling approximately 200 design rules. The physical verification software used is a commercial software whose function library contains approximately 300 geometric operation functions.

[0043] like Figure 1 As shown, a layout design rule generation method based on multimodal intelligent agent technology includes the following steps:

[0044] S1. Convert the source file of the design rule manual containing text, tables and images into a structured text format and divide it according to the process level;

[0045] S2. Use a visual model to extract semantics from the schematic diagrams in the segmented file to generate image description text; fuse the image description text with the original rule text, with the original rule text as the priority and the image description text as supplementary information, to obtain an enhanced rule description.

[0046] S3. Based on the enhanced rule description, a set of candidate functions is selected from the function library through vector similarity retrieval, and combined with the historical experience rule library, the first-level reasoning model is used to determine the function or combination of functions required to generate the current rule.

[0047] S4. Query the function usage database to obtain the usage information of the function or function combination determined by the decision, and retrieve the auxiliary constraints and global descriptions of the process level involved in the current rule from the design rule manual, and retrieve the relevant code specifications from the experience knowledge base. All the retrieved information is used as knowledge enhancement and input into the second-level reasoning model to guide it to reason according to the preset thinking chain, and finally generate executable layout design rule check code.

[0048] It should be noted that this invention is applied to the scenario of automatically generating layout design rule check code in the field of integrated circuit design automation. Specifically, when an integrated circuit design team or process manufacturer obtains a new design rule manual, it needs to convert the hundreds or even thousands of geometric constraints, electrical constraints, hierarchical relationships, and other natural language and image descriptions in the manual into rule check code executable by specific physical verification software. Traditional manual conversion methods are time-consuming, error-prone, and highly dependent on expert experience. The concept of this invention is to construct a multimodal intelligent agent system that works collaboratively with a visual model and a reasoning model. First, the design rule manual is structurally decomposed. Then, the visual model extracts key geometric information from the schematic diagram and merges it with the text description. Next, a two-level reasoning model architecture is used to realize function decision-making and code generation. Knowledge enhancement is achieved by leveraging a function usage database, a process-level global knowledge base, and a historical experience knowledge base. Finally, the model is guided to complete the output of executable code according to a preset thought chain.

[0049] In this embodiment, in step S1, the source file of the design rule manual is first obtained. This file is in PDF format and contains various modal information such as natural language paragraphs, parameter tables, geometric diagrams, and annotation boxes.

[0050] It should be noted that directly inputting this PDF file into a large model would result in excessively long context, and the model would be unable to simultaneously understand the image content. Therefore, this step uses the open-source document parsing tool MinerU to convert the PDF file. MinerU can recognize text blocks, table structures, and image regions in a PDF and output them as Markdown files. It also stores each image separately as a PNG or JPEG file, marking it with its image index path within the Markdown text. The conversion accuracy of MinerU directly affects the quality of subsequent information extraction; this embodiment uses its default configuration and enables table recognition and image extraction functions.

[0051] Furthermore, after the conversion is complete, a Markdown file containing the complete manual content and an image folder are obtained. To further reduce the context length of a single processing step and to prevent the model from being interfered with by other irrelevant process layer information when processing a certain rule, this embodiment segments the Markdown file according to the process layer.

[0052] The specific approach is as follows: Predefine a list of process-level keywords, such as NW, ACT, BPR, GCON, GCUT, DUMMY, etc. Scan the Markdown file; when these keywords are detected as chapter headings, extract all content from the current heading up to the next level heading and save it as a separate Markdown subfile. Each subfile is named "layer_name_drm.md", and the index of images referenced within the subfile is retained. For content not explicitly belonging to a specific level but belonging to global descriptions, such as the scope of design rules, unit conventions, and layer mapping tables, save it separately as a global file "global_drm.md".

[0053] Through the above segmentation, the original 300-page manual was split into 18 process layer files and 1 global file, with each file's length controlled between 5 and 20 pages, effectively controlling the context length of subsequent model inputs.

[0054] In this embodiment, step S2 addresses the deficiency that plain text descriptions cannot fully convey visual information such as measurement directions, prohibition relationships, and relative positions in geometric schematic diagrams. Specifically, for each process layer file segmented in step S1, the images within are processed one by one. This includes the following sub-steps:

[0055] S21. Legend Mapping Construction and Rule Candidate Set Establishment: For a process layer file, all referenced images are read. For each image, the visual model Qwen-VL-Instruct is called with the prompt: "Please identify the process layer names represented by different colors or patterns in the image and output a color-to-layer name mapping table." The visual model returns a mapping result similar to "Red: GCON, Blue: GCUT, Green: DUMMY." A common legend mapping table is maintained. If the same process layer has inconsistent colors in different images, the mapping with the most frequent occurrence is used, and a conflict alarm is recorded. Simultaneously, all rule names contained in that layer are extracted from the table of the process layer file. For example, from Table 5: GCON Rules, 18 rule names from GCON.1 to GCON.18 are extracted to form the rule candidate set for that layer. This candidate set is used to subsequently associate image information with specific rules.

[0056] S22. Rule-level Graph Semantic Extraction: For each rule in the candidate set, search for nearby image references in its corresponding process layer file. If a reference exists, use the image file, the constructed legend mapping table, and the rule name as input to initiate a question-and-answer session with the visual model. This embodiment designs the following standardized question-and-answer template:

[0057] Question 1: Please describe which process layers are involved in the measurement object in the figure? Represent them using the layer names in the mapping table.

[0058] Question 2: Is the measurement method side-to-side, corner-to-corner, or something else? Please specify the measurement direction (horizontal / vertical / oblique).

[0059] Question 3: Are there any red crosses or other prohibited symbols in the diagram? If so, indicate the prohibited contact or overlap.

[0060] Question 4: Are there any formulas or values ​​labeled in the image? If so, please extract them completely.

[0061] Question 5: Are there any exception areas (e.g., hollowed-out areas, restricted areas)? If so, please describe their locations.

[0062] Furthermore, the visual model answers each question based on the image content. A diagram illustrating the GCON.9 rule is shown below. Figure 2 As shown in the left figure, the visual model returns the following: the layers involved are GCON and GCUT; the measurement method is edge-to-edge vertical measurement, and the measurement object is the vertical distance between the bottom edge of the GCON region and the top edge of the GCUT region; there is no cross indicating prohibition; there are no formulas; there are no exception areas.

[0063] For rule GCON.14, the diagram is as follows: Figure 2 As shown in the right figure, the visual model returns the following: the layers involved are GCON, GCUT, and DUMMY; the measurement method is contact prohibition; there are black crosses at the junction of GCON and DUMMY, and also at the junction of GCON and GCUT, indicating that direct contact is prohibited; there are no formulas; there are no exception areas.

[0064] The image description text above is stored in JSON format, with each rule corresponding to a description object.

[0065] S23. Multimodal conflict resolution and semantic fusion: Since schematic diagrams may be drawn in idealized shapes to emphasize certain geometric relationships, such as simplifying polygons into rectangles, or visual models may produce misinterpretations, it is necessary to set conflict resolution rules.

[0066] This embodiment employs the following priority strategy: the natural language description in the original rule text has the highest priority. For example, the text description of GCON.9 is "Minimum vertical spacing between GCON and GCUT," which explicitly includes "vertical spacing" and "minimum," serving as an unshakeable constraint. Information extracted by the visual model that contains explicit numerical values ​​or formulas, such as spacing > 0.0135 μm, is considered a high-confidence supplement; information containing only directional descriptions, such as up / down or left / right directions, is considered a secondary priority reference. When the text description conflicts with the visual information, for example, the text describes horizontal spacing while the visual description describes vertical spacing, the system prioritizes the text and records the conflict in the log. Finally, the original rule text is concatenated with the priority-filtered image description text to form the enhanced rule description.

[0067] For example, for GCON.9, the enhancements are described as follows:

[0068] Rule Name: GCON.9; Original Description: Minimum vertical spacing between GCON and GCUT; Image Supplement: The measurement objects are the bottom edge of GCON and the top edge of GCUT, and the measurement direction is vertical edge to edge, with no prohibition relationship.

[0069] In this embodiment, step S3 addresses the context burden problem caused by the large size of the function library, which cannot all be input into the model. Through two-level filtering and a single model decision, only the few functions most relevant to the current rule are provided to the inference model. This includes the following sub-steps:

[0070] S31. Construct a candidate set of functions from the description semantics: This sub-step involves two parallel retrieval channels.

[0071] The first channel: Function retrieval based on TF-IDF and cosine similarity; a function definition library is pre-built, where each function includes a function name, function description text (extracted from the software manual), and a parameter list example. For the enhanced description text of the current rule, normalized word segmentation is first performed: English stop words are removed, uppercase is converted to lowercase, and verb and noun stems are extracted.

[0072] For example, after enhanced descriptive word segmentation in GCON.9, the keyword set {minimum, vertical, spacing, between, GCON, GCUT, bottom, edge, top, edge, edge-to-edge} is obtained. Similarly, the same word segmentation process is performed on the description text of each function in the function library. Then, a term frequency-inverse document frequency (TF) vector is constructed. The TF calculation uses a log-normalized formula:

[0073] ;in, For terms t In the document d The number of times it appears in the text.

[0074] The formula for calculating Inverse Document Frequency (IDF) is:

[0075] ;in, N This represents the total number of functions in the function library. For included terms t The number of function documents.

[0076] Ultimately, the vector representation of each document is the TF-IDF value of each term. The rule description vector is denoted as... Function document vector is denoted as The formula for calculating cosine similarity is: .

[0077] After calculating the similarity between the rule description and each function document, the functions are sorted from highest to lowest similarity, and the top 15 functions and their similarity scores are added to the candidate function set. C 1. For GCON.9, the search results show that the similarity for the space function is 0.32, for the geom_spacing function it is 0.28, for the edge_separation function it is 0.21, and for other functions it is below 0.20.

[0078] The second approach involves similarity retrieval based on a historical experience rule base. An experience knowledge base is pre-built, storing DRC rule codes that have been validated in past projects. To eliminate differences at specific process layers, process layer names are anonymized during storage: specific layer names, such as GCON, are replaced with LAYER_A, and GCUT with LAYER_B, while preserving the structural characteristics of the rules. Each experience record includes the anonymized rule description text, the function combination used, and the generated code snippet.

[0079] For the enhanced description of the current rule, anonymization is also performed: the specific layer name in the rule is replaced with a generic placeholder. Then, using the same TF-IDF and cosine similarity methods as the first channel, the similarity between the current rule description and each record in the empirical knowledge base is calculated. The five empirical rules with the highest similarity and their function combinations are selected as enhanced information and added to the candidate set. C 2. For GCON.9, a similar rule was found: the vertical spacing measurement between LAYER_A and LAYER_B, which used the space function and constructed a horizontal derivative layer. This rule is used as an example.

[0080] merge C 1 and C 2. Obtain the total set of candidate functions. C . C The function may contain duplicate functions; after deduplication, there are usually 18-20 functions.

[0081] S32, Inference Model Decision: The enhanced rule description generated in step S2 and the candidate function set obtained in step S31 are combined. C The names and brief usage descriptions of each function, along with the three most relevant historical experience examples, are combined into a structured prompt, which is then input into the first-level inference model, DeepSeek-V3.1.

[0082] For example, the first-level inference model outputs a decision; for GCON.9, the model outputs: the decision uses the space function. The rationale is that the rule explicitly requires measuring the minimum distance between perpendicular edges, and the space function supports direction parameters and is consistent with historical experience. For GCON.14, the candidate function search results show that geom_net_interact has the highest similarity of 0.28, but after inference, the model believes that geom_interact better fits the semantics of contact prohibition because geom_interact can detect whether polygons in two layers have overlapping areas or shared edges, while geom_net_interact is mainly used for network connectivity checks. The final decision uses the geom_interact function.

[0083] In this embodiment, step S4 ensures the accuracy and executability of the output code through multi-source knowledge enhancement and a structured thought chain. It includes the following sub-steps:

[0084] S41. Enhanced Function Usage Knowledge: A pre-built SQLite database is used to structure and store the function usage manuals for the physical verification software. Specifically, each function entry in the manual is parsed to extract fields such as function name, function signature, parameter list, parameter type, return value, usage example, and precautions. For example, for the `space` function, the database records include: function name `space`, signature `space(layer1, layer2, constraint)`, parameter descriptions `layer1` and `layer2` are geometric layers, `constraint` is an inequality such as `<0.0135`, and example `space( M1 M2 < 0.1)`, etc.

[0085] Furthermore, after determining the function, the function name is used as the key to query the SQLite database to obtain complete usage information for that function. For GCON.9, the space function is queried to obtain its detailed usage. For GCON.14, the geom_interact function is queried to obtain its usage format: geom_interact(layer1, layer2 [, options]), where options can be omitted.

[0086] S42. Supplementing knowledge based on the manual content and past experience: This sub-step includes three aspects.

[0087] Firstly, the process-level auxiliary constraint retrieval involves parsing the rule names to extract the relevant process levels. For example, GCON.9 involves GCON and GCUT, while GCON.14 involves GCON, GCUT, and DUMMY. Then, auxiliary constraints related to that level are retrieved throughout the entire manual, including global files. For instance, retrieving GCON might yield rules such as GCON layers must not have acute angles and GCON minimum width is 0.02 μm. Although these constraints are not directly required by the current rules, they must be followed when generating derived layers or performing geometric operations. The retrieval method uses regular expressions to match layer names and extracts contextual paragraphs. The retrieved auxiliary constraints are then compiled into supplementary knowledge.

[0088] Secondly, global specification knowledge: The pre-saved global_drm.md file is read, which contains global specifications from the design rule manual, such as unit system (μm), coordinate system, default precision, and inter-layer priority. These global specifications are used as fixed knowledge input.

[0089] Thirdly, regarding the retrieval of the experience knowledge base: For complex rules, such as those involving latch-up effects or centerline distance constraints, the general large model lacks relevant process knowledge. In this embodiment, the experience knowledge base is stored using a vector database. Each record contains a rule type label, a list of keywords, anonymized code templates, and expert comments. For GCON.9, the keywords "vertical spacing" and "derived layer" are retrieved, yielding the experience knowledge that when measuring vertical distance, it is usually necessary to first construct a horizontal derived layer. The construction method is: Derived layer name = ORIG_LAYER_HE, defined as the intersection of the original layer and the horizontal auxiliary line. The specific code is: GCON_HE = geom_and( GCON, HORIZONTAL_LINE ). This knowledge is used as a supplement.

[0090] S43. Reasoning Paradigm Based on Thinking Chains: To prevent unpredictable errors from arising in black-box reasoning of the model, such as... Figure 3 As shown, a mandatory thought chain was designed, requiring the second-level reasoning model to output the thought results of each step in sequence, finally generating code. The prompt explicitly requires the model to follow the following thought chain:

[0091] The first step is to understand the semantics of the rules and restate the constraints: The model needs to re-describe the core requirements of the rules in its own language, and clarify the measurement object, measurement type (spacing, enclosure, contact, area, etc.), direction, and numerical threshold.

[0092] The second step involves the selection and understanding of the summary information and the experience knowledge base: the model needs to indicate which image information is used and which is discarded, and explain the reasons; at the same time, it needs to explain which practices are borrowed from the experience knowledge base.

[0093] The third step is to recommend the use of candidate functions: the model needs to confirm whether the function used in step S3 is appropriate, and if it is not appropriate, adjustments can be proposed.

[0094] The fourth step is to supplement the understanding of conceptual information: the model needs to explain the impact of process-level auxiliary constraints on the current rules.

[0095] Step 5: Reasons for selecting or discarding parameters of the function: The model needs to list all parameters of the selected function and explain the value of each parameter and why it is chosen. For the space function, the parameters are layer1, layer2, and constraint. Model output: layer1 is GCON_HE (the horizontally derived GCON layer) because the bottom edge needs to be measured; layer2 is GCUT_HE (the horizontally derived GCUT layer) because the top edge needs to be measured; constraint is <0.0135 because the rule requires a minimum vertical spacing, i.e., a measured value less than 0.0135 is a violation.

[0096] Step 6, Derivative Layer Reuse: The model needs to specify whether derived layers are needed, and how they are defined. For GCON.9, the model output requires two derived layers: GCON_HE and GCUT_HE. GCON_HE is defined as the intersection of the GCON layer and an infinitely long horizontal line at its bottom edge. This can be achieved by shrinking the GCON layer by half its width using the geom_size function and then taking the bottom edge. Since there is no direct bottom edge function in the software library, a method from the knowledge base is used: first, copy the GCON layer, and then use geom_inside to filter out the part that intersects with the horizontal auxiliary line.

[0097] Step 7: Self-check for common problems before output: The model needs to be checked to see if the units are correct, if the values ​​include units, if the layer names are spelled correctly, if the brackets match, and if the comments are added, etc.

[0098] After completing the above thought process, the model outputs the final code.

[0099] For example, for GCON.9, the generated code is:

[0100] RULE GCON.9 {

[0101] ;Minimum vertical spacing between GCON and GCUT

[0102] space (GCON_HE GCUT_HE < 0.0135)

[0103] }

[0104] It should be noted that this code checks the minimum vertical spacing between the GCON and GCUT layers in the layout, using the `space` function for spacing measurement. To accurately measure the vertical distance, the original GCON and GCUT layers are not used directly; instead, their horizontal edge-derived layers, GCON_HE and GCUT_HE, are used. A parameter <0.0135 in the function indicates a constraint: the spatial distance between these two derived layer graphics must be greater than 0.0135 μm. If the distance is less than this value, it will be marked as a violation of design rules.

[0105] For example, for GCON.14, the generated code is as follows:

[0106] RULE GCON.14 {

[0107] ;GCON may not interact with GCUT or DUMMY

[0108] geom_interact(GCON GCUT)

[0109] geom_interact(GCON DUMMY)

[0110] }

[0111] It's important to note that this code checks that the GCON layer must not touch or overlap with the GCUT or DUMMY layers. It uses the `geom_interact` function to examine the interaction between geometries. This function selects shapes that share a portion of their area with another layer's polygons, completely overlap, or have edges touching. The rule contains two statements, one checking for prohibited interactions between the GCON and GCUT layers, and the other checking for interactions between the GCON and DUMMY layers. If such interactions exist, the corresponding shape area is marked as a violation.

[0112] Example 2

[0113] Figure 4 A schematic diagram of an electronic device structure that can be used to implement Embodiment 1 of the present invention is shown. The electronic device 10 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (such as helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0114] like Figure 4 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory 12 or a random access memory 13, which is communicatively connected to the at least one processor 11. The memory stores a computer program that can be executed by the at least one processor 11, and the computer program is executed by the at least one processor 11 to enable the at least one processor 11 to perform the method provided by the present invention.

[0115] Furthermore, the processor 11 can perform various appropriate actions and processes based on the computer program stored in the read-only memory 12 or the computer program loaded from the storage unit 18 into the random access memory 13. The random access memory 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, the read-only memory 12, and the random access memory 13 are interconnected via a bus 14. The I / O interface 15 is also connected to the bus 14.

[0116] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0117] Furthermore, processor 11 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as methods for resource management of a database.

[0118] In some specific embodiments, the method for managing database resources can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via read-only memory 12 and / or communication unit 19. When the computer program is loaded into random access memory 13 and executed by processor 11, one or more steps of the method for managing database resources described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to perform the method for managing database resources by any other suitable means (e.g., by means of firmware).

[0119] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard parts (ASSPs), systems-on-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0120] Computer programs used to implement the methods of the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The computer programs can be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0121] In the context of this invention, a computer-readable storage medium stores computer instructions that, when executed by a processor, implement the method for resource management of a database provided by this invention. The computer-readable storage medium may be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. The computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, the computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory, read-only memory 12, erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0122] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a cathode ray tube (CRT)) or a liquid crystal display (LCD monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0123] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0124] Optionally, the computing system may include clients and servers. Clients and servers are generally geographically separated and typically interact via a communication network. The client-server relationship is established by computer programs running on the respective computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system. This addresses the shortcomings of traditional physical hosts and Virtual Private Server (VPS) services, such as high management difficulty and weak business scalability.

[0125] The above description is based on the preferred embodiments of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered exemplary and non-limiting in all respects. The scope of the invention is defined by the appended claims rather than the foregoing description, and all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0126] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for design rule generation based on multi-modal agent technology, characterized in that, Includes the following steps: S1. Convert the source file of the design rule manual containing text, tables and images into a structured text format and divide it according to the process level; S2. Use a visual model to extract semantics from the schematic diagrams in the segmented file to generate image description text; fuse the image description text with the original rule text, with the original rule text as the priority and the image description text as supplementary information, to obtain an enhanced rule description. S3. Based on the enhanced rule description, a set of candidate functions is selected from the function library through vector similarity retrieval, and combined with the historical experience rule library, the first-level reasoning model is used to determine the function or combination of functions required to generate the current rule. S4. Query the function usage database to obtain the usage information of the function or function combination determined by the decision, and retrieve the auxiliary constraints and global descriptions of the process level involved in the current rule from the design rule manual, and retrieve the relevant code specifications from the experience knowledge base. All the retrieved information is used as knowledge enhancement and input into the second-level reasoning model to guide it to reason according to the preset thinking chain, and finally generate executable layout design rule check code.

2. The layout design rule generation method based on multimodal intelligent agent technology according to claim 1, characterized in that, Step S2 includes the following sub-steps: S21. For images in the same process layer, maintain a mapping table between legend colors and process layer names, and extract the rule names contained in the layer from the document table to establish a rule candidate set. S22. Input the legend mapping, image file and rule name into the visual model, and extract the effective information in the diagram through question and answer. The effective information includes: involved layer, measurement object, measurement method, prohibition symbol and formula expression. S23. When there is a semantic conflict between the rule text description and the diagram information, the rule text description shall be given first priority, the quantitative relationship in the diagram shall be given as a highly reliable supplement, and the directional relationship in the diagram shall be given as the second priority, and the reasoning model shall make the choice.

3. The layout design rule generation method based on multimodal intelligent agent technology according to claim 1, characterized in that, Step S3 includes the following sub-steps: S31. On the one hand, based on TF-IDF vectors and cosine similarity, the similarity between the rule description text and the function names and function descriptions in the function library is calculated to obtain the first candidate function set; On the other hand, the rule description text is matched with the anonymized rule description in the historical experience knowledge base, and the empirical rules with the highest similarity and their function combinations are extracted as the second candidate function set. S32. Merge the first candidate function set and the second candidate function set, and pass the names of the functions, usage descriptions and relevant historical experience examples as input to the first-level inference model, which will then decide on the final required function or combination of functions.

4. The layout design rule generation method based on multimodal intelligent agent technology according to claim 1, characterized in that, In step S4, the function usage database is constructed in the following way: the function usage manual of the physical verification software is preprocessed, indexed by function name, and the function usage instructions, format specifications and parameter descriptions are used as values ​​to create an SQLite database.

5. The layout design rule generation method based on multimodal intelligent agent technology according to claim 1, characterized in that, In step S4, the knowledge enhancement further includes: extracting the process level involved by parsing the name of the current rule, retrieving auxiliary constraint information related to that level in the entire document, and retrieving global description information from the design rule manual, and inputting the above information as supplementary knowledge into the second-level reasoning model.

6. The layout design rule generation method based on multimodal intelligent agent technology according to claim 1, characterized in that, In step S4, the construction and retrieval of the experience knowledge base includes: checking code files according to existing layout design rules, and constructing the knowledge base after anonymizing specific process levels; during retrieval, the knowledge base is constructed by calculating the vector similarity of the rule description text or by keyword matching.

7. The layout design rule generation method based on multimodal intelligent agent technology according to claim 1, characterized in that, In step S4, the preset thought chain includes the following reasoning steps executed in sequence: rule semantic understanding and constraint target restatement, selection of image summary information and empirical knowledge, applicability analysis of candidate functions, understanding of supplementary concept information, analysis of the selection or rejection of function parameters one by one, judgment of the reuse of derived layers, and self-check of common problems before output.

8. The layout design rule generation method based on multimodal intelligent agent technology according to claim 1, characterized in that, Steps S3 and S4 employ a two-level inference model architecture: the first-level inference model receives the simplified candidate function information and uses it to decide on the required function or combination of functions; the second-level inference model receives the decision result from the first level and queries to obtain detailed function usage information and various domain knowledge for the final code generation.

9. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to at least one of the processors; The memory stores a computer program that is executed by at least one of the processors, such that the at least one processor is able to execute the layout design rule generation method based on multimodal intelligent agent technology as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the layout design rule generation method based on multimodal intelligent agent technology as described in any one of claims 1-8.

Images