Two-dimensional and three-dimensional computer aided design dual-path collaborative method based on large model

Abstract

The application discloses a two-dimensional and three-dimensional computer-aided design double-path cooperation method based on a large model, and relates to the technical field of intelligent design. The method comprises the following steps: receiving natural language design requirement data and identifying corresponding output dimension types; constructing a prompt word configured with a unified white list constraint based on the data, calling a large language model to output a structured design specification with a unified format; distributing and analyzing the structured design specification according to the output dimension types, obtaining a vector drawing file via a two-dimensional processing flow, and obtaining a parameterized geometric model via a three-dimensional processing flow; respectively carrying out visual conformity evaluation and topological consistency checking on the two types of achievements, obtaining multi-modal feedback data, constructing feedback prompt backfilling a large language model, incrementally modifying the structured design specification, and obtaining final computer-aided design data. The scheme is beneficial to improving the standardization, cooperation and iterative stability of CAD intelligent generation.

Description

Technical Field

[0001] This application relates to the field of intelligent design technology, and in particular to a dual-path collaborative method for two-dimensional and three-dimensional computer-aided design based on a large model. Background Technology

[0002] Computer-aided design (CAD) is a core supporting technology for modern industrial digital R&D and intelligent manufacturing. It is widely used in fields such as machinery manufacturing, architectural engineering, and industrial product design, and is a crucial foundation for realizing digital product modeling, drawing output, and process iteration. With the continuous iteration and updating of large-scale modeling technology, the automatic conversion of natural language design requirements into CAD data based on large-scale language models has become an important development direction in the field of intelligent design, effectively improving the problems of high barriers to entry and low iteration efficiency associated with traditional manual modeling and drawing.

[0003] Current intelligent CAD generation solutions mostly adopt a single-dimensional, independent generation architecture, isolating the 2D drawing generation and 3D model building links. They lack unified standards for requirement analysis and specification output, making it impossible to achieve collaborative 2D and 3D generation based on the same design requirements. Furthermore, existing technologies lack a unified constraint mechanism to standardize the output content of large models. Free text output from large models easily exceeds the processing boundaries of downstream CAD analysis and modeling, frequently causing problems such as analysis anomalies, missing drawing information, and modeling execution errors. In addition, existing generation systems are single-loop output structures, lacking a multimodal feedback mechanism adapted to 2D visual evaluation and 3D topology verification. 2D visual evaluation results and 3D topology verification results cannot form standardized feedback signals for iterative transmission. When design results deviate, only manual readjustment of requirement parameters is possible, failing to achieve automatic correction and continuous convergence of design specifications. Overall, the level of intelligence and design stability cannot meet the high-precision, batch, and automated CAD design requirements of industry. Summary of the Invention

[0004] The purpose of this application is to provide a dual-path collaborative method for two-dimensional and three-dimensional computer-aided design based on a large model, which aims to solve the technical problems of fragmented two-dimensional and three-dimensional design links and lack of closed-loop iteration capability in intelligent generation in the prior art.

[0005] To achieve the above objectives, this application provides a dual-path collaborative method for two-dimensional and three-dimensional computer-aided design based on a large model, comprising: Receive natural language design requirement data and identify the output dimension type corresponding to the natural language design requirement data; Based on the natural language design requirements data, construct prompt words with unified whitelist constraints, and call the large language model to output structured design specifications with unified format; The structured design specifications are parsed using a dual-path distribution based on the output dimension type, resulting in vector paper files through a two-dimensional processing flow and parametric geometric models through a three-dimensional processing flow. The visual conformity of the vector paper file is evaluated, the topological consistency of the parametric geometric model is verified, and multimodal feedback data is obtained. Based on the multimodal feedback data, feedback prompts are constructed and backfilled into the large language model. The structured design specifications are then incrementally modified to obtain the final computer-aided design data.

[0006] In some optional embodiments, receiving natural language design requirement data and identifying the output dimension type corresponding to the natural language design requirement data includes: Collect raw text information input by the user, perform noise filtering and sentence normalization on the raw text information, and generate natural language design requirement data; Load a preset dimension recognition lexicon, extract entities from the natural language design requirement data, and obtain a set of design dimension keywords; Match the design dimension keyword set with the preset dimension recognition lexicon to generate a dimension confidence score. The output dimension type corresponding to the natural language design requirement data is determined based on the dimension confidence value.

[0007] In some optional embodiments, prompt words with a unified whitelist constraint are constructed based on the natural language design requirement data, and a structured design specification with a unified output format is called from the large language model, including: Retrieve the pre-stored set of unified whitelist constraint rules; The natural language design requirement data is used to extract parameters to obtain a set of requirement parameters; The prompt words are generated by combining the set of required parameters with the set of unified whitelist constraint rules. Input the prompt words into the large language model and obtain the original design text output by the large language model; The original design text is structurally transformed according to a preset standardized template to obtain structured design specifications.

[0008] In some optional embodiments, the prompt words are assembled by combining the set of requirement parameters with the set of unified whitelist constraint rules, including: Iterate through the set of required parameters and verify whether each parameter matches the set of unified whitelist constraint rules. Filter to obtain a subset of valid requirement parameters that pass the validation; The subset of valid demand parameters and the set of unified whitelist constraint rules are encapsulated in text to generate the main body of the prompt words; Attach format constraint instructions to the body of the prompt word to form a complete prompt word.

[0009] In some optional embodiments, the original design text is structurally transformed according to a preset standardized template to obtain structured design specifications, including: The original design text is parsed and segmented to separate several independent design information fragments; Match each of the aforementioned design information fragments with the field identifiers defined within the preset standardized template; Fill the design information fragments into the corresponding field identifier positions to generate intermediate specification data; Perform field integrity checks on the intermediate specification data to obtain the structured design specifications.

[0010] In some optional embodiments, the structured design specification is parsed using a dual-path distribution based on the output dimension type, resulting in a vector drawing file via a two-dimensional processing flow and a parametric geometric model via a three-dimensional processing flow, including: Based on the output dimension type, the structured design specifications are divided into two-dimensional design information and three-dimensional design information. The two-dimensional design information is distributed to the two-dimensional processing flow, and the three-dimensional design information is distributed to the three-dimensional processing flow. The two-dimensional processing flow generates vector drawing files, and the three-dimensional processing flow generates parametric geometric models.

[0011] In some optional embodiments, running the two-dimensional processing flow to generate vector paper files includes: The primitive information of the two-dimensional design information is extracted to obtain a set of basic drawing units; Load preset two-dimensional drawing rules, and perform position mapping and attribute assignment on the basic drawing unit set; The basic drawing unit set, after being assigned values ​​according to the vector file encoding rules, is used to obtain the vector drawing file.

[0012] In some optional embodiments, running the 3D processing flow to generate a parametric geometric model includes: Extract a set of parametric modeling instructions from the 3D design information; The set of parametric modeling instructions is sent into a preset isolated operating environment; The parametric modeling instruction set is executed sequentially within the isolated operating environment to obtain a parametric geometric model.

[0013] In some optional embodiments, visual compliance evaluation is performed on the vector paper file, topological consistency verification is performed on the parametric geometric model, and multimodal feedback data is obtained, including: Extract the primitive features and annotation information from the vector paper file to generate a two-dimensional visual sample; The visual conformity of the two-dimensional visual sample is evaluated based on a preset visual evaluation standard to obtain a two-dimensional evaluation result. Traverse the geometric edges, vertices, and constraint relationships of the parametric geometric model to construct a topological association dataset; The topological consistency of the topologically associated dataset is verified according to the preset topological determination rules to obtain the three-dimensional verification result. By integrating the two-dimensional evaluation results with the three-dimensional verification results, multimodal feedback data is constructed.

[0014] In some optional embodiments, feedback prompts are constructed based on the multimodal feedback data and backfilled into the large language model to incrementally modify the structured design specifications, resulting in final computer-aided design data, including: Analyze the multimodal feedback data to identify design fields with anomalies; Associate the design field identifier with the corresponding field within the structured design specification to generate a field correction instruction; Integrate the field correction instructions and constraint text to construct feedback prompts; The feedback prompt is input into the large language model, and the specification correction information output by the large language model is received; The structured design specifications are incrementally modified based on the specification correction information to obtain the final computer-aided design data.

[0015] The above-mentioned technical solution of this application has at least the following beneficial technical effects: The technical solution of this application achieves standardized and unified parsing of design requirements by uniformly identifying the output dimension types of natural language design requirement data, thus avoiding the fragmented processing of two-dimensional and three-dimensional requirements. By configuring prompts with a unified whitelist constraint to standardize the output of the large language model, it obtains a uniformly formatted structured design specification, effectively constraining the output boundary of the large model and avoiding parsing anomalies and design omissions caused by free text output. Through dual-path distribution and parsing of the structured design specification, vector drawing files and parametric geometric models are generated simultaneously via two-dimensional and three-dimensional processing flows, enabling single-requirement-driven collaborative generation of two-dimensional and three-dimensional designs. By simultaneously conducting visual compliance assessment and topological consistency verification and obtaining multimodal feedback data, it uniformly collects two-dimensional and three-dimensional design quality deviation information. By constructing feedback prompts from the multimodal feedback data to fill the large language model, it incrementally modifies the structured design specification and outputs the final computer-aided design data, constructing a complete closed-loop iterative system of automatic machine verification, automatic feedback, and automatic correction. This system overcomes the limitations of traditional single-loop generation and reliance on manual error correction, significantly improving the standardization, collaboration, and iterative stability of intelligent CAD generation. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating an embodiment of the dual-path collaborative method for two-dimensional and three-dimensional computer-aided design based on a large model provided in this application. Figure 2 This is a system architecture diagram of the unified whitelist structured output and two-dimensional and three-dimensional dual-path collaborative parsing provided in this application; Figure 3 This is a multimodal feedback closed-loop timing diagram of the CAD intelligent generation and feedback method provided in this application; Figure 4 This is a measured output image of the vector graphic generated by two-dimensional processing path parsing of the ceramic teapot provided in this application; Figure 5 This is a measured output image of the parametric geometric model of the ceramic teapot provided in this application, generated after being processed through a 3D processing path and executed in a secure sandbox. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to specific embodiments and accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of this application. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessary confusion regarding the concepts of this application.

[0018] The embodiments described in this application are only some, not all, of the embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments described herein without inventive effort are within the scope of protection of this application.

[0019] CAD is a core supporting technology for modern industrial digital R&D and intelligent manufacturing. It is widely used in fields such as machinery manufacturing, architectural engineering, and industrial product design, and is an important foundation for realizing digital product modeling, drawing output, and process iteration. With the continuous iteration and updating of large model technology, the automatic conversion of natural language design requirements into CAD data based on large language models has become an important development direction in the field of intelligent design, effectively improving the problems of high threshold and low iteration efficiency of traditional manual modeling and drawing.

[0020] Current intelligent CAD generation solutions mostly adopt a single-dimensional, independent generation architecture, isolating the 2D drawing generation and 3D model building links. They lack unified standards for requirement analysis and specification output, making it impossible to achieve collaborative 2D and 3D generation based on the same design requirements. Furthermore, existing technologies lack a unified constraint mechanism to standardize the output content of large models. Free text output from large models easily exceeds the processing boundaries of downstream CAD analysis and modeling, frequently causing problems such as analysis anomalies, missing drawing information, and modeling execution errors. In addition, existing generation systems are single-loop output structures, lacking a multimodal feedback mechanism adapted to 2D visual evaluation and 3D topology verification. 2D visual evaluation results and 3D topology verification results cannot form standardized feedback signals for iterative transmission. When design results deviate, only manual readjustment of requirement parameters is possible, failing to achieve automatic correction and continuous convergence of design specifications. Overall, the level of intelligence and design stability cannot meet the high-precision, batch, and automated CAD design requirements of industry.

[0021] To address the aforementioned technical issues, this application provides a dual-path collaborative method for two-dimensional and three-dimensional computer-aided design based on a large model.

[0022] The technical solution of this application achieves standardized and unified parsing of design requirements by uniformly identifying the output dimension types of natural language design requirement data, thus avoiding the fragmented processing of two-dimensional and three-dimensional requirements. By configuring prompts with a unified whitelist constraint, it standardizes the output of the large language model, obtaining structured design specifications with a unified format, effectively constraining the output boundaries of the large model, and avoiding parsing anomalies and design omissions caused by free text output. Through dual-path distribution and parsing of the structured design specifications, vector drawing files and parametric geometric models are generated simultaneously via two-dimensional and three-dimensional processing flows, achieving single-requirement-driven collaborative generation of two-dimensional and three-dimensional designs. By simultaneously conducting visual compliance assessment and topological consistency verification and obtaining multimodal feedback data, it uniformly collects two-dimensional and three-dimensional design quality deviation information. By constructing feedback prompts from the multimodal feedback data and backfilling them into the large language model, it incrementally modifies the structured design specifications and outputs the final computer-aided design data, constructing a complete closed-loop iterative system of automatic machine verification, automatic feedback, and automatic correction. This system overcomes the limitations of traditional single-loop generation and reliance on manual error correction, significantly improving the standardization, collaboration, and iterative stability of intelligent CAD generation.

[0023] In some alternative embodiments, please refer to Figure 1 The dual-path collaborative method for two-dimensional and three-dimensional computer-aided design based on a large model includes the following steps: Step S1: Receive natural language design requirement data and identify the output dimension type corresponding to the natural language design requirement data; In some optional embodiments, step S1 includes the following specific steps: Step S11: Collect the raw text information input by the user, perform noise filtering and sentence normalization on the raw text information, and generate natural language design requirement data. Specifically, the system collects raw text information input by the user, performs noise filtering and sentence normalization on the raw text information to generate the natural language design requirement data; performs requirement intent classification, identifies the target object category, delivery format, and detail elements, and distinguishes the output dimension type as two-dimensional drawing type and three-dimensional model type. For example, if the input is a two-dimensional drawing requirement for a teapot, it is determined to be a two-dimensional drawing type; if the input is a three-dimensional modeling requirement, it is determined to be a three-dimensional model type.

[0024] Step S12: Load the preset dimension recognition lexicon, extract entities from the natural language design requirement data, and obtain a set of design dimension keywords; Specifically, a preset dimension recognition lexicon is loaded, entity extraction is performed on the natural language design requirement data, and keywords for entity, size, and output type are extracted and aggregated to form a set of design dimension keywords.

[0025] Step S13: Match the set of design dimension keywords with the preset dimension recognition thesaurus to generate dimension confidence scores; Specifically, the set of design dimension keywords is matched one by one with the preset dimension recognition word library, and the number of matching hits is counted to generate a dimension confidence value in the range of 0 to 1.

[0026] Step S14: Determine the output dimension type corresponding to the natural language design requirement data based on the dimension confidence score.

[0027] Specifically, the output dimension type corresponding to the natural language design requirement data is determined based on the dimension confidence value threshold; the output dimension type is divided into two-dimensional drawing type and three-dimensional model type, and parallel distribution of the two types can be supported.

[0028] Step S2: Based on the natural language design requirement data, construct prompt words with a unified whitelist constraint, and call the large language model to output a structured design specification with a unified format; In some optional embodiments, step S2 includes the following specific steps: Step S21: Retrieve the pre-stored unified whitelist constraint rule set; Specifically, a pre-stored unified whitelist constraint rule set is retrieved. This unified whitelist constraint rule set includes a two-dimensional entity whitelist, a two-dimensional annotation whitelist, a three-dimensional application programming interface (API) whitelist, and a list of disabled APIs. The two-dimensional entity whitelist includes circles, line segments, arcs, lightweight polylines, polylines, text, ellipses, circular fills, and polygon fills. The two-dimensional annotation whitelist includes linear annotations, aligned annotations, radius annotations, diameter annotations, and angle annotations. The three-dimensional API whitelist includes cube primitives, cylinder primitives, sphere primitives, working plane switching, coordinate translation, rotation transformation, union operation, difference operation, fillet processing, and chamfer processing. The list of disabled APIs includes solids of revolution, lofting, sweeping, spline curves, elliptical arcs, closures, wireframes, and moving to.

[0029] Step S22: Extract parameters from the natural language design requirement data to obtain a set of requirement parameters; Specifically, parameter fields are extracted from the natural language design requirement data, separating size parameters, material parameters, and geometric construction parameters, and summarizing them to obtain a set of requirement parameters.

[0030] Step S23: Combine the set of requirement parameters with the set of unified whitelist constraint rules to assemble the prompt words; In some optional embodiments, step S23 further includes the following specific steps: Step S231: Traverse the set of requirement parameters and verify whether each parameter matches the unified whitelist constraint rule set; Specifically, all parameter items in the set of required parameters are traversed one by one, and the entity and modeling interface corresponding to the parameter are verified to match the unified whitelist constraint rule set. Parameters that hit the list of disabled application interfaces are intercepted.

[0031] Step S232: Filter to obtain a subset of valid requirement parameters that have passed the verification; Specifically, invalid parameters that fail the verification are removed, and a subset of valid requirement parameters that fully match the whitelist rules are selected and retained.

[0032] Step S233: Encapsulate the subset of valid requirement parameters and the unified whitelist constraint rule set into text to generate the main body of the prompt words; Specifically, the subset of valid requirement parameters and the complete unified whitelist constraint rule set are encapsulated in structured text and written into the system prompt word template to generate the prompt word body; the two-dimensional drawing type is matched with the two-dimensional dedicated prompt word template and injected with the two-dimensional whitelist and the two-dimensional domain few sample example set; the three-dimensional model type is matched with the three-dimensional dedicated prompt word template and injected with the three-dimensional whitelist, the list of disabled interfaces, the component intersection contract, the component visibility contract and the three-dimensional domain few sample example set.

[0033] Step S234: Attach format constraint instructions to the prompt body to form a complete prompt.

[0034] Specifically, a structured JSON output format constraint instruction and a minimum field integrity requirement instruction are appended to the end of the prompt word body to form a complete prompt word constrained by a unified whitelist.

[0035] Step S24: Input prompt words into the large language model and obtain the original design text output by the large language model; Specifically, the complete prompt words are fed into a privately deployed domestic general-purpose open-source large language model (Tongyi Qianwen, Shendu Qiusuo, Zhipu, Shusheng Puyu, Doubao, etc., and industrial fine-tuning large model) for inference, and the unorganized original design text is received from the model output.

[0036] Step S25: Perform a structured transformation on the original design text according to the preset standardized template to obtain the structured design specifications.

[0037] In some optional embodiments, step S25 further includes the following specific steps: Step S251: Perform text segmentation and parsing on the original design text to separate it into several independent design information fragments; Specifically, the original design text is parsed by text segmentation, cut by field delimiters, and separated into independent design information fragments such as requirements, parameters, geometry, layers, and scripts.

[0038] Step S252: Match each design information fragment with the field identifiers defined within the preset standardized template; Specifically, each of the independent design information fragments is matched one by one with the predefined field identifiers of the preset standardized template. The preset standardized template includes requirement analysis field, parameter field, standard reference field, geometric planning field, dimension planning field, verification point field, layer field, title bar field, drawing instruction field / parametric modeling script field.

[0039] Step S253: Fill the design information fragments into the corresponding field identifier positions to generate intermediate specification data; Specifically, the matched design information fragments are filled into the corresponding field identifier positions in the template to generate unvalidated intermediate specification data.

[0040] Step S254: Perform field integrity checks on the intermediate specification data to obtain the structured design specifications.

[0041] Specifically, the intermediate specification data undergoes a completeness check of all required fields. Missing fields are automatically filled with default template values. After the check passes, a structured CAD specification in a unified JSON format is output, which is the structured design specification.

[0042] Step S3: Based on the output dimension type, perform dual-path distribution parsing of the structured design specifications, obtain vector graphics files through a two-dimensional processing flow, and obtain parametric geometric models through a three-dimensional processing flow. In some optional embodiments, step S3 includes the following specific steps: Step S31: Divide the two-dimensional design information and three-dimensional design information within the structured design specification based on the output dimension type; Specifically, based on the output dimension type determined in S14, the structured design specifications are split as follows: drawing instruction fields, layer fields, and title bar fields are extracted as two-dimensional design information; parametric modeling script fields and three-dimensional dimension parameters are extracted as three-dimensional design information; and the two types of information are split and processed in parallel.

[0043] Step S32: Distribute the 2D design information to the 2D processing flow, and distribute the 3D design information to the 3D processing flow; Specifically, the two-dimensional design information is distributed to the two-dimensional hierarchical parsing module, and the three-dimensional design information is distributed to the three-dimensional security sandbox execution module.

[0044] Step S33: Run the 2D processing flow to generate vector drawing files, and run the 3D processing flow to generate parametric geometric models.

[0045] In some optional embodiments, a two-dimensional processing flow is run to generate a vector drawing file, including the following specific steps: extracting primitive information from the two-dimensional design information to obtain a set of basic drawing units; loading preset two-dimensional drawing rules to perform position mapping and attribute assignment on the set of basic drawing units; and organizing the assigned set of basic drawing units according to vector file encoding rules to obtain a vector drawing file.

[0046] Specifically, all graphic elements are extracted from the drawing instruction field in the two-dimensional design information. The entity type of each graphic element is identified and the corresponding layer attributes are bound. The layers include an outline layer, a centerline layer, a dimension layer, and a text layer. Entity drawing is performed layer by layer. Chinese character set font styles are injected into text entities, and dimension coverage parameters with forced integer display are configured for all dimension entities. The complete title bar is drawn, and after generating a vector paper file, it is converted into a visual image file and a portable document file for visual evaluation.

[0047] In some optional embodiments, the three-dimensional processing flow generates a parametric geometric model, including the following specific steps: extracting a set of parametric modeling instructions from the three-dimensional design information; sending the set of parametric modeling instructions into a preset isolated running environment; and sequentially executing the set of parametric modeling instructions within the isolated running environment to obtain the parametric geometric model.

[0048] Specifically, complete parametric modeling script fields are extracted from the 3D design information, and the script undergoes a three-layer security sandbox process: the first layer is static verification of the abstract syntax tree to intercept dangerous syntax; the second layer is module import whitelist verification to restrict third-party modules; and the third layer is restricted execution of built-in functions to isolate sensitive host resources. After the sandbox verification passes, the script is executed to obtain the parametric geometric model object. A four-view visualization image is generated using a multi-angle rendering method, and the isometric view, front view, top view, and side view are output side by side in the same image, along with the title block information within the specifications.

[0049] Step S4: Perform visual compliance assessment on vector graphics files, perform topological consistency verification on parametric geometric models, and obtain multimodal feedback data; In some optional embodiments, step S4 includes the following specific steps: Step S41: Extract the primitive features and annotation information from the vector paper file to generate a two-dimensional visual sample; Specifically, the vector paper files output by S33 are rendered into visual images, and all primitive outlines, dimension annotations, and text labels within the images are extracted, integrated to generate two-dimensional visual samples, and then sent to a privately deployed visual language model.

[0050] Step S42: Based on the preset visual evaluation criteria, evaluate the visual conformity of the two-dimensional visual samples to obtain the two-dimensional evaluation results; Specifically, based on preset visual evaluation criteria, the visual big language model performs visual compliance evaluation on the two-dimensional visual samples and outputs a two-dimensional evaluation result that includes compliance score field, satisfied requirement field, identified problem field, and improvement suggestion field.

[0051] Step S43: Traverse the geometric edges, vertices, and constraint relationships of the parametric geometric model to construct a topological association dataset; Specifically, the modeling engine interface is called to traverse all geometric edges, vertices, and constraint relationships of the parametric geometric model, extract all independent geometric bodies and count their number, calculate the upper and lower bounding boxes of the X / Y / Z axes in the orthogonal coordinate system of each independent geometric body, and construct a complete topological association dataset.

[0052] Step S44: Perform topology consistency verification on the topology association dataset according to the preset topology judgment rules to obtain the three-dimensional verification result; Specifically, the topology association dataset is verified according to the preset topology judgment rules: if the number of independent geometric objects is equal to one, the topology is determined to be consistent; if the number is greater than one, the topology is determined to be abnormal. The number of geometric objects and the spatial distribution information of each bounding box are integrated into the three-dimensional verification result.

[0053] Step S45: Integrate the two-dimensional evaluation results and the three-dimensional verification results to construct multimodal feedback data.

[0054] Specifically, the two-dimensional evaluation results and three-dimensional verification results are spliced ​​together and uniformly packaged into a standardized machine-readable text structure to construct complete multimodal feedback data.

[0055] Step S5: Construct feedback prompts based on multimodal feedback data and backfill them into the large language model, incrementally modify the structured design specifications, and obtain the final computer-aided design data.

[0056] In some optional embodiments, step S5 includes the following specific steps: Step S51: Analyze the multimodal feedback data and locate the design field identifiers where anomalies exist; Specifically, the multimodal feedback data is parsed to extract problem items and topology anomalies, and the corresponding design field identifiers of deviations within the structured design specifications are located in reverse.

[0057] Step S52: Associate the design field identifier with the corresponding field within the structured design specification, and generate a field correction instruction; Specifically, the abnormal field identifiers are associated one by one with the original fields of the structured design specifications, generating field correction instructions that only modify the deviation fields.

[0058] Step S53: Integrate field correction instructions and constraint text construction feedback prompts; Specifically, a dedicated feedback prompt template is matched according to the output dimension type. For two-dimensional models, a visual feedback template is used to fill in drawing problems and improvement suggestions, while for three-dimensional models, a topological feedback template is used to fill in the number of geometric objects and the distribution information of bounding boxes. The feedback prompt is constructed by integrating field correction instructions and whitelist invariant constraint text, and the minimum necessary modification constraints are injected, requiring the large language model to only modify the feedback-related fields, while the remaining fields remain unchanged.

[0059] Step S54: Input the feedback prompts into the large language model and receive the specification correction information output by the large language model; Specifically, the feedback prompts indicating completion and the previous version of the complete structured design specifications are added to the context input of the large language model, triggering model inference and receiving specification correction information that is only incrementally corrected by the model output.

[0060] Step S55: Incrementally modify the structured design specifications based on the specification correction information to obtain the final computer-aided design data.

[0061] Specifically, the original structured design specifications are partially incrementally replaced using specification correction information without altering the unbiased fields; the S3 to S5 iteration process is executed repeatedly until the visual score meets the standard, the topology verification passes, or the preset maximum number of dialogue rounds is reached; after the iteration terminates, two-dimensional vector exchange files, raster images, PDF files, three-dimensional stereolithography files, and parametric standard exchange files are exported; all exported files, the final structured design specifications, and the evaluation report are packaged into a compressed file and delivered as the final complete computer-aided design data.

[0062] The technical solution of this application will be described in detail below with reference to a specific embodiment.

[0063] In this embodiment, the user inputs natural language design requirements: to generate two-dimensional engineering drawings and a three-dimensional parametric model of a ceramic teapot. The system then executes a dual-path collaborative method for two-dimensional and three-dimensional computer-aided design based on a large model.

[0064] Please see Figure 2 , Figure 2 The system architecture of unified whitelist structured output and two-dimensional and three-dimensional dual-path collaborative parsing is shown. The entire system consists of a requirement parsing module, a large language model inference module, a structured CAD specification storage unit, a two-dimensional path parsing module, a three-dimensional path parsing module, a visual large language model evaluation module, a topology consistency verification module, and an incremental modification backfilling module. Each module works together in sequence to form a complete closed-loop iterative link.

[0065] First, the requirements analysis module is activated to collect the raw text information input by the user. The raw text information is then subjected to noise filtering and sentence normalization to generate natural language design requirements data. A preset dimension recognition lexicon is loaded to extract entities and obtain a set of design dimension keywords. The keyword set is matched with the lexicon to generate dimension confidence scores. Based on the confidence scores, the output dimension type is determined. Simultaneously, parallel processing of two-dimensional drawings and three-dimensional models is supported.

[0066] Subsequently, the large language model inference module is invoked to retrieve a unified whitelist constraint rule set, which includes a two-dimensional entity whitelist, a two-dimensional annotation whitelist, a three-dimensional application programming interface (API) whitelist, and a list of disabled APIs. Parameters are extracted from the natural language design requirement data to obtain a requirement parameter set. All parameters are traversed and verified, and a subset of valid requirement parameters is selected. This subset is then encapsulated with the whitelist rules to generate the prompt word body, and format constraint instructions are appended to form a complete prompt word. The prompt word is then fed into the privately deployed general-purpose open-source large language model for inference to obtain the original design text. The original design text is then segmented, parsed, and its fields matched and populated to obtain intermediate specification data. After completing field integrity verification, the structured design specifications are output.

[0067] Please see Figure 3 , Figure 3 The multimodal feedback closed-loop timing flow of the CAD intelligent generation and feedback method is shown. The entire operation process is divided into initial generation rounds and incremental modification rounds. The system splits the structured design specifications according to the output dimension type, separates the two-dimensional design information and the three-dimensional design information, and distributes them to two independent processing links respectively.

[0068] The 2D path parsing module receives 2D design information, completes layered parsing, configures Chinese fonts and integer annotation rules, extracts primitive information, and generates vector paper files.

[0069] Please see Figure 4 , Figure 4 This diagram shows the measured output of a vector graphic of a ceramic teapot generated through 2D processing path parsing. The graphic fully depicts the structural outlines of the teapot body, handle, spout, lid, and knob, and indicates the material as ceramic and the scale as 1:1, meeting the engineering 2D drawing delivery specifications. The system renders the vector graphic into a visual image and sends it to the Visual Language Model Evaluation Module. The Visual Language Model Evaluation Module performs a visual compliance assessment and outputs a 2D evaluation result including a score, problem items, and improvement suggestions. In this embodiment, the first round of evaluation results in a score of 8, with missing dimension annotations.

[0070] The 3D path parsing module receives 3D design information, extracts a set of parametric modeling instructions, and sends them into a three-layer security sandbox environment. It then sequentially performs static verification of the abstract syntax tree, module import whitelist verification, and built-in function resource isolation. After the verification is passed, it executes the modeling instructions to generate a parametric geometric model.

[0071] Please see Figure 5 , Figure 5 The image shows the measured output of the parametric geometric model of the ceramic teapot generated after the 3D processing path is executed in a safe sandbox. The model synchronously outputs isometric views, front view, top view, and side view, displaying the effect from multiple angles. The label fully records the external dimensions, material, version, and drawing number information, intuitively presenting the 3D solid structure of the teapot. The topology consistency verification module traverses the model vertices, edges, and constraint relationships to construct a topology association dataset. It is determined that the number of independent geometric bodies is 1, and the topology verification result is normal.

[0072] The incremental modification backfill module integrates two-dimensional evaluation results and three-dimensional verification results to construct multimodal feedback data, parses the data to locate abnormal field identifiers within the structured design specifications, and generates field correction instructions; it matches feedback prompt templates, constructs feedback prompts to backfill into the general open-source large language model, performs local incremental modifications on the original structured design specifications, and generates a new version of the structured design specifications with supplemented dimension annotations.

[0073] The system uses the updated structured design specifications to perform dual-path distribution parsing again, completes a new round of drawing and model generation, and the secondary visual evaluation score is 9, reaching the preset qualified threshold. The topology verification continues to pass the state, and the closed-loop iteration process is terminated.

[0074] Finally, the incremental modification backfill module packages and outputs vector paper exchange files (DXF, PDF), 3D model (STL, STEP), and complete process document compressed packages according to preset formats, forming the final computer-aided design data delivered to the user.

[0075] This application aims to protect a dual-path collaborative method for 2D and 3D computer-aided design based on a large model. This technical solution achieves standardized and unified parsing of design requirements by uniformly identifying the output dimension type of natural language design requirement data, thus avoiding the fragmented processing problem of splitting 2D and 3D requirements. By configuring prompts with a unified whitelist constraint to standardize the output of the large language model, a uniformly formatted structured design specification is obtained, effectively constraining the output boundary of the large model and avoiding parsing anomalies and design omissions caused by free text output. Furthermore, by performing dual-path distribution and parsing of the structured design specification, vectors are simultaneously generated through 2D and 3D processing flows. Drawing files and parametric geometric models enable collaborative generation of 2D and 3D designs driven by single requirements. By simultaneously conducting visual compliance assessments and topological consistency checks and acquiring multimodal feedback data, 2D and 3D design quality deviation information is uniformly collected. By constructing a feedback prompt backfilling large language model from the multimodal feedback data, the structured design specifications are incrementally modified and the final computer-aided design data is output. This constructs a complete closed-loop iterative system of automatic machine verification, automatic feedback, and automatic correction, breaking away from the limitations of traditional single open-loop generation and reliance on manual error correction, and significantly improving the standardization, collaboration, and iterative stability of CAD intelligent generation.

[0076] It should be understood that the specific embodiments described above are merely illustrative or explanatory of the principles of this application and do not constitute a limitation thereof. Therefore, any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of this application should be included within the protection scope of this application. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

Claims

1. A dual-path collaborative method for two-dimensional and three-dimensional computer-aided design based on a large model, characterized in that, include: Receive natural language design requirement data and identify the output dimension type corresponding to the natural language design requirement data; Based on the natural language design requirements data, construct prompt words with unified whitelist constraints, and call the large language model to output structured design specifications with unified format; The structured design specifications are parsed using a dual-path distribution based on the output dimension type, resulting in vector paper files through a two-dimensional processing flow and parametric geometric models through a three-dimensional processing flow. The visual conformity of the vector paper file is evaluated, the topological consistency of the parametric geometric model is verified, and multimodal feedback data is obtained. Based on the multimodal feedback data, feedback prompts are constructed and backfilled into the large language model. The structured design specifications are then incrementally modified to obtain the final computer-aided design data.

2. The dual-path collaborative method for two-dimensional and three-dimensional computer-aided design based on a large model according to claim 1, characterized in that, Receive natural language design requirement data, identify the output dimension type corresponding to the natural language design requirement data, including: Collect raw text information input by the user, perform noise filtering and sentence normalization on the raw text information, and generate natural language design requirement data; Load a preset dimension recognition lexicon, extract entities from the natural language design requirement data, and obtain a set of design dimension keywords; Match the design dimension keyword set with the preset dimension recognition lexicon to generate a dimension confidence score. The output dimension type corresponding to the natural language design requirement data is determined based on the dimension confidence value.

3. The dual-path collaborative method for two-dimensional and three-dimensional computer-aided design based on a large model according to claim 1, characterized in that, Based on the aforementioned natural language design requirements data, a unified whitelist constraint is configured for prompt words. The large language model is then invoked to output a structured design specification with a unified format, including: Retrieve the pre-stored set of unified whitelist constraint rules; The natural language design requirement data is used to extract parameters to obtain a set of requirement parameters; The prompt words are generated by combining the set of required parameters with the set of unified whitelist constraint rules. Input the prompt word into the large language model and obtain the original design text output by the large language model; The original design text is structurally transformed according to a preset standardized template to obtain structured design specifications.

4. The dual-path collaborative method for two-dimensional and three-dimensional computer-aided design based on a large model according to claim 3, characterized in that, The prompt words are assembled by combining the set of required parameters with the set of unified whitelist constraint rules, including: Iterate through the set of required parameters and verify whether each parameter matches the set of unified whitelist constraint rules. Filter to obtain a subset of valid requirement parameters that pass the validation; The subset of valid demand parameters and the set of unified whitelist constraint rules are encapsulated in text to generate the main body of the prompt words; Attach format constraint instructions to the body of the prompt word to form a complete prompt word.

5. The dual-path collaborative method for two-dimensional and three-dimensional computer-aided design based on a large model according to claim 3, characterized in that, The original design text is structurally transformed according to a preset standardized template to obtain structured design specifications, including: The original design text is parsed and segmented to separate several independent design information fragments; Match each of the aforementioned design information fragments with the field identifiers defined within the preset standardized template; Fill the design information fragments into the corresponding field identifier positions to generate intermediate specification data; Perform field integrity checks on the intermediate specification data to obtain the structured design specifications.

6. The dual-path collaborative method for two-dimensional and three-dimensional computer-aided design based on a large model according to claim 1, characterized in that, The structured design specifications are parsed using a dual-path distribution method based on the output dimension type. A vector drawing file is obtained via a two-dimensional processing flow, and a parametric geometric model is obtained via a three-dimensional processing flow, including: Based on the output dimension type, the structured design specifications are divided into two-dimensional design information and three-dimensional design information. The two-dimensional design information is distributed to the two-dimensional processing flow, and the three-dimensional design information is distributed to the three-dimensional processing flow. The two-dimensional processing flow generates vector drawing files, and the three-dimensional processing flow generates parametric geometric models.

7. The dual-path collaborative method for two-dimensional and three-dimensional computer-aided design based on a large model according to claim 6, characterized in that, Running the two-dimensional processing flow to generate vector paper files includes: The primitive information of the two-dimensional design information is extracted to obtain a set of basic drawing units; Load preset two-dimensional drawing rules, and perform position mapping and attribute assignment on the basic drawing unit set; The basic drawing unit set, after being assigned values ​​according to the vector file encoding rules, is used to obtain the vector drawing file.

8. The dual-path collaborative method for two-dimensional and three-dimensional computer-aided design based on a large model according to claim 6, characterized in that, Running the 3D processing flow to generate a parametric geometric model includes: Extract a set of parametric modeling instructions from the 3D design information; The set of parametric modeling instructions is sent into a preset isolated operating environment; The parametric modeling instruction set is executed sequentially within the isolated operating environment to obtain a parametric geometric model.

9. The dual-path collaborative method for two-dimensional and three-dimensional computer-aided design based on a large model according to claim 1, characterized in that, The vector graphics file is subjected to visual compliance evaluation, the parametric geometric model is subjected to topological consistency verification, and multimodal feedback data is obtained, including: Extract the primitive features and annotation information from the vector paper file to generate a two-dimensional visual sample; The visual conformity of the two-dimensional visual sample is evaluated based on a preset visual evaluation standard to obtain a two-dimensional evaluation result. Traverse the geometric edges, vertices, and constraint relationships of the parametric geometric model to construct a topological association dataset; The topological consistency of the topologically associated dataset is verified according to the preset topological determination rules to obtain the three-dimensional verification result. By integrating the two-dimensional evaluation results with the three-dimensional verification results, multimodal feedback data is constructed.

10. The dual-path collaborative method for two-dimensional and three-dimensional computer-aided design based on a large model according to claim 1, characterized in that, Based on the multimodal feedback data, feedback prompts are constructed and backfilled into the large language model. The structured design specifications are then incrementally modified to obtain the final computer-aided design data, including: Analyze the multimodal feedback data to identify design fields with anomalies; Associate the design field identifier with the corresponding field within the structured design specification to generate a field correction instruction; Integrate the field correction instructions and constraint text to construct feedback prompts; The feedback prompt is input into the large language model, and the specification correction information output by the large language model is received; The structured design specifications are incrementally modified based on the specification correction information to obtain the final computer-aided design data.

Images