A 3D modeling simulation method and system for nuclear fuel performance analysis

Abstract

The application relates to a three-dimensional modeling simulation method, system and equipment for nuclear fuel performance analysis, and relates to the fields of artificial intelligence and nuclear physics. The method comprises the following steps: based on adaptive probes and multi-source consistency verification, parameters are extracted from multi-modal heterogeneous documents to construct a physical parameter knowledge base; based on multi-agent collaboration, parameters are called from the knowledge base to generate a three-dimensional modeling simulation script, the script is executed in a safe sandbox, and iteration is performed based on physical constraints; multi-modal intermediate state data of a simulation process are extracted to construct visual-physical evolution memory, and dynamic confidence pruning is performed; a knowledge graph is constructed based on the knowledge base and the memory, parameter sparse areas are identified to autonomously generate and explore new tasks. The application overcomes the defects of traditional modeling, such as artificial dependence on high models, multi-modal extraction illusion, long cycle trial and error forgetting, and passive response, and realizes high-fidelity automatic simulation.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a three-dimensional modeling and simulation method and system for nuclear fuel performance analysis. Background Technology

[0002] With the evolution of large language models and multi-agent systems, artificial intelligence-driven scientific research paradigms are playing a vital role in various fields. In the field of nuclear engineering, nuclear fuel performance analysis and three-dimensional thermo-mechanical coupling simulation, as core components in assessing reactor safety, possess extremely high physical complexity and engineering rigor.

[0003] Traditional nuclear fuel performance analysis largely relies on conventional deterministic numerical simulation programs and pre-defined empirical models, such as FRAPCON / FRAPTRAN. These programs typically depend on pre-defined empirical formulas and simplified one- or two-dimensional meshes, with fixed input interfaces. Researchers must manually extract characteristic parameters and write input files in specific formats. To address the challenge of high-fidelity three-dimensional modeling, advanced tools such as BISON (based on the MOOSE framework) and the large-scale commercial finite element analysis software Abaqus are widely used. However, engineers still need to use CAE interfaces or write complex Python scripts to complete geometric modeling and mesh generation.

[0004] In terms of acquiring the input parameters required for high-fidelity simulation, early industrial document digitization mainly relied on optical character recognition and rule-based regular expression extraction. In recent years, multimodal pre-trained document analysis models, represented by LayoutLMv3 and Donut, have become mainstream, but they are essentially still perception models based on static pattern recognition and lack deep logical reasoning.

[0005] To lower the simulation threshold, some technical solutions attempt to directly generate simulation code using large language models, or to achieve task decomposition and multi-step reasoning through multi-agent collaborative systems such as LangChain and AutoGen. However, most of these systems follow a passive response mode driven by human instructions, and their memory modules only support static text storage.

[0006] While existing technologies have provided some assistance in industrial simulation and document processing, the following significant shortcomings still exist in demanding scenarios involving nuclear fuel performance analysis and high-fidelity three-dimensional thermo-mechanical coupling modeling: (1) Traditional modeling paradigms rely heavily on manual data extraction and writing of complex scripts, which can easily lead to computational deadlock in high-dimensional parameter spaces, and engineering experience is difficult to accumulate. (2) Existing multimodal extraction technologies (such as OCR or general VLM) lack the ability to sniff out features and reason logically for unlabeled raw data. When processing heterogeneous documents, they are prone to producing factual "illusions" and cross-modal feature fragmentation, and cannot guarantee the high fidelity of parameters. (3) Existing agent collaborative modeling mostly adopts passive one-way generation and lacks a safe sandbox verification closed loop based on underlying physical laws (such as nuclear physics constraints); and existing memory modules mostly only support static plain text, directly discarding the "visual-physical" intermediate debugging trajectory, which contains grid distortion and other highly valuable engineering features, causing the system to fall into catastrophic forgetting and high repeated trial and error costs in complex modeling. (4) Existing AI simulation systems are completely limited by the passive response mode of "human command-driven", lacking a self-planning active cognitive link, making it difficult to exhaust all potential critical states under multi-physical field coupling, and easily overlooking the hidden physical boundaries under extreme working conditions. Summary of the Invention

[0007] This application provides a 3D modeling and simulation method and system for nuclear fuel performance analysis. It aims to address the problems of existing technologies, such as the lack of deep logical reasoning and adaptive probe mechanisms when analyzing massive amounts of unstructured and highly heterogeneous industrial or nuclear engineering archives, which easily leads to factual "illusions" and fragmented cross-modal knowledge during the extraction of key multimodal features. It also addresses the issues of traditional high-fidelity 3D modeling in nuclear physics heavily relying on manually written scripts, and the lack of a secure sandbox verification loop based on underlying physical laws in existing AI simulation frameworks, which easily leads to computational deadlock and low fault tolerance when facing complex multi-physics strongly coupled problems. Furthermore, it addresses the problem that existing multi-agent collaborative modeling mechanisms only support single static text memory, directly discarding highly valuable multimodal intermediate visual debugging trajectories, resulting in catastrophic forgetting and high repeated trial-and-error costs during long-cycle complex modeling. Finally, it addresses the problem that existing AI-assisted simulation frameworks are completely limited by a passive response mode driven by human commands, lacking autonomous cognition and proactive planning, making it difficult to exhaust all potential critical states under multi-physics coupling, and easily overlooking high-dimensional implicit physical boundaries and extreme conditions.

[0008] Firstly, this application provides a 3D modeling and simulation method for nuclear fuel performance analysis. The method operates within a multi-agent system architecture and includes: S1. Obtain multimodal heterogeneous documents containing nuclear fuel experimental data. Based on the on-demand triggering strategy and dynamic code probe, perform multimodal knowledge extraction and multi-source consistency verification on the multimodal heterogeneous documents to obtain high-fidelity standard triples. Construct a structured physical parameter knowledge base based on the standard triples. S2, obtain simulation task instructions in natural language form, parse the instructions based on the multi-agent collaborative mechanism and retrieve the corresponding parameters from the physical parameter knowledge base to generate a 3D modeling simulation script; S3, based on the simulation script, set the running parameters of the safety sandbox, collect the physical state results and multimodal intermediate state data output by the solver in real time, if an error occurs, perform error analysis and correction on the error information based on physical rules, and iteratively correct the 3D modeling simulation script until execution is successful; S4 compares the intermediate state data of successful iteration steps and failed iteration steps, extracts the core modification actions that lead to the reversal of the simulation state through difference analysis, encapsulates them into multimodal experience tuples containing visual-physical features, and stores them in the evolutionary memory bank for retrieval and reuse in subsequent modeling tasks. S5. Construct a knowledge graph based on the physical parameter knowledge base and the evolutionary memory base, scan the high-dimensional parameter space of the knowledge graph to identify sparse knowledge regions, and autonomously generate and execute verification tasks under extreme conditions based on the identification results.

[0009] Secondly, this application also provides a 3D modeling and simulation system for nuclear fuel performance analysis. The system includes: The knowledge extraction module is used to parse and process multimodal heterogeneous documents, convert unstructured data into structured physical parameters, and build a parameter knowledge base. The multi-agent modeling module is used to automatically transform simulation tasks into executable 3D modeling and simulation processes within a multi-agent collaborative framework. The simulation execution module is used to execute simulation tasks in a controlled environment and collect running status and result data in real time. The memory management module is used to store, organize, and evolve multimodal data during the simulation process, enabling experience accumulation and reuse. The autonomous exploration module is used to proactively expand the simulation task space and discover potential critical operating conditions without human intervention.

[0010] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in any of the above-mentioned embodiments.

[0011] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any of the preceding claims.

[0012] The aforementioned 3D modeling and simulation method and system for nuclear fuel performance analysis has the following advantages compared with existing technologies: First, by triggering probes on demand and verifying consistency from multiple sources, the barrier to extracting multimodal heterogeneous data is broken, effectively suppressing the "illusion" of large models and ensuring the high fidelity of the input simulation model data; Secondly, through the closed loop of intelligent agents with secure sandbox execution and reflection on underlying physical rules, it can autonomously analyze the root cause of errors and iteratively repair the code, achieving a high degree of automation in complex multi-physics coupling modeling, and significantly reducing modeling time and trial-and-error costs. Furthermore, by incorporating intermediate visual tuning trajectories into memory and establishing an elimination mechanism through "visual-physical" multimodal evolutionary memory and dynamic confidence pruning, this study is the first to completely solve the catastrophic forgetting problem under long-term operation of complex systems. Finally, through multi-dimensional knowledge graph topology scanning and autonomous task generation, the traditional "passive instruction-driven" paradigm is broken, giving the system the proactive cognitive ability to explore unknown physical boundaries without human intervention.

[0013] This application's solution achieves a complete closed loop from data extraction to intelligent modeling, and then to memory storage and autonomous exploration, significantly improving the intelligence level and engineering efficiency of nuclear fuel performance analysis. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is an application environment diagram of a three-dimensional modeling and simulation method for nuclear fuel performance analysis in one embodiment; Figure 2 This is a flowchart illustrating a three-dimensional modeling and simulation method for nuclear fuel performance analysis in one embodiment. Figure 3 This is a schematic diagram of multimodal heterogeneous document knowledge extraction based on on-demand triggering and adaptive probes in one embodiment; Figure 4 This is a schematic diagram of 3D modeling and simulation verification based on multi-agent collaboration and underlying physical reflection in one embodiment; Figure 5 This is a schematic diagram of multimodal evolutionary memory deposition and dynamic confidence pruning in one embodiment; Figure 6 This is a schematic diagram illustrating the exploration of breaking through the boundaries of instruction-driven autonomous cognition in one embodiment; Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0017] This application provides a 3D modeling and simulation method for nuclear fuel performance analysis, which can be applied to, for example... Figure 1 The application environment shown.

[0018] In one embodiment, such as Figure 2 As shown, a 3D modeling and simulation method for nuclear fuel performance analysis is provided. Taking the application of this method to a server as an example, the method includes the following steps: S1: Multimodal knowledge extraction and multi-source consistency verification based on on-demand triggering strategy and dynamic code probe.

[0019] One possible design involves multimodal knowledge extraction based on an on-demand triggering strategy and dynamic code probes, including: By using a document parsing model, the text and tables in the multimodal heterogeneous documents are decoupled from their layout and restructured to obtain a unified Markdown text format. For the image data in the multimodal heterogeneous document, a global semantic summary of the image data is extracted using a large visual language model. The global semantic summary is then concatenated with the original image caption to construct a cross-modal placeholder tag, which is then embedded in the Markdown text format in situ. Upon receiving the target parameter extraction instruction, a candidate context subset containing the target parameter is obtained from the Markdown text format based on a dual-path retrieval strategy; if the candidate context subset contains the cross-modal placeholder label and the numerical information is missing, the visual language big model is triggered to extract local numerical features from the original high-resolution image cropped block. For unlabeled power history files, the header sequence of the power history file is extracted as metadata prior. A prompt word template is constructed by combining it with the entity name of the target parameter. An abstract syntax tree parsing script is dynamically generated using a large language model. The parsing script is executed in a security sandbox to output the specific value of the target parameter.

[0020] In one possible design, the dual-path retrieval strategy for obtaining a subset of candidate contexts containing the target parameter in the Markdown text format includes: The BM25 algorithm is used to calculate the keyword matching score between the target parameter extraction instruction and the Markdown text format segmentation block; The instructions and text blocks are transformed into dense vectors using a domain-fine-tuned embedding model, and the cosine similarity score between them is calculated. The keyword matching score and the cosine similarity score are weighted and fused, and the Top-K text blocks are recalled as candidate context subsets based on the ranking result of the fused scores.

[0021] In one possible design, step S1, the multi-source consistency verification, includes: For multiple candidate observations of the same physical parameter extracted in different modes, their units are uniformly converted to the preset standard dimensions to obtain a normalized numerical sequence. Calculate the variance of the normalized numerical sequence. If the variance is lower than a preset engineering tolerance threshold, then take the mean as the standard value. If the variance is greater than or equal to the engineering tolerance threshold, the softmax average probability of the information source modality token of each observation is calculated as the prior confidence weight, and the average value is taken after removing low confidence noise points with weights lower than the preset value.

[0022] In the embodiments of this application, such as Figure 3 As shown, firstly, the open-source document parsing model MinerU is used to decouple the layout of unstructured documents, converting text and tables into a unified machine-readable Markdown format. For images, VLM is used to extract global semantic summaries. and the original image caption Concatenate and build cross-modal placeholder tags that are embedded in place within Markdown.

[0023] Secondly, upon receiving the nuclear physics parameter extraction command, a dual-path recall strategy using BM25 and cosine similarity is employed to obtain the context. If the recalled context contains placeholders and the values ​​are insufficient, VLM is triggered to perform local deep understanding of the cropped blocks of the original high-resolution image.

[0024] Next, a dynamic code probe is used to parse unlabeled power history files: the header sequence of the power history file is extracted as metadata prior, and combined with the physical definition of the target parameter (such as average fuel consumption), a Prompt template is constructed as input to a large language model. The large model infers implicit column mapping relationships and dynamically generates a Python abstract syntax tree parsing script. This script performs I / O reads and timing integral calculations in a secure sandbox and outputs the scalar value of the target parameter.

[0025] Finally, multi-source consistency verification: for samples extracted from the same parameter... Each candidate observation is uniformly converted to standard dimensions. The variance of the normalized numerical sequence is calculated. If the variance is lower than the engineering tolerance threshold, the mean is taken. For example, geometric parameters such as "shell outer diameter" are set to 0.5%, and thermodynamic evolution parameters such as "fuel consumption" are set to 5.0%. If there is a conflict, the softmax average probability of the parameter token output by the large model is calculated as the prior confidence weight. Noise points with too low weights are removed. Finally, high-fidelity parameter triples are output and written into the structured physical parameter knowledge base.

[0026] S2: Generates 3D modeling and simulation scripts based on multi-agent collaboration and executes them in a sandbox.

[0027] In the embodiments of this application, such as Figure 4 As shown, the system receives natural language simulation task instructions and activates the multi-agent collaborative mechanism: Planning agent: Decomposes instructions into a sequence of subtasks in a directed acyclic graph. If no matching operator is available in the toolkit, a stream of coded instructions for "new operator development" is issued.

[0028] Encoded intelligent agent: Retrieves specific values ​​and units from a structured knowledge base, enforces type checks and geometric topological constraints, and generates high-precision Abaqus finite element simulation execution scripts.

[0029] The execution agent runs the script by calling the underlying finite element solver within a secure sandbox with resource isolation, such as Docker. During execution, it monitors the solver's intermediate state logs, convergence residual curves, and potential error messages such as mesh distortion in real time, generating an execution status feedback set.

[0030] The reflective agent is used to receive the execution state feedback set and perform root cause analysis by combining the built-in nuclear physics prior rules and solver error semantics.

[0031] The reflective agent generates a reflective report containing specific suggestions for modifying parameters or adjusting the grid. The reflection report is sent back to the coding agent, which then drives the coding agent to regenerate the 3D modeling simulation script for the next iteration, until the simulation physical state results satisfy the physical conservation law.

[0032] S3: Iterative correction based on physical rules.

[0033] In the embodiments of this application, such as Figure 4 As shown, the reflexive agent acts as a review node. If the executing agent reports an execution failure, the reflexive agent performs root cause analysis based on built-in nuclear physics prior rules and generates a reflexive report. The reflexive report is then fed back to the encoding agent to drive iterative code rewriting, forming a closed loop of "generation-reflection-encapsulation" for error correction without human intervention.

[0034] S4: Extract multimodal intermediate data to construct visual-physical evolutionary memory and dynamically prune it.

[0035] In the embodiments of this application, such as Figure 5 As shown, trajectory distillation is triggered after the simulation finally converges. The code AST differences between failed and successful iteration steps are aligned, and core modification actions are automatically extracted. This is then encapsulated into a multimodal empirical tuple. ,in This is the text for an error message scenario. These are intermediate visual features such as grid distortion. To correct the rules, the data is stored in the evolutionary memory. The dynamic pruning mechanism of the memory is as follows: Confidence weights are initialized for each empirical tuple. In subsequent reuse, if this experience leads to simulation convergence, a positive reward is given for updating. ( (Set to 0.1); if this results in two errors or is not called 10 times consecutively, a penalty will be imposed. ( Set to 0.3). Periodically scan the memory bank; when a certain empirical tuple... Below the threshold When this happens, it is automatically pruned and removed, effectively mitigating the catastrophic forgetting that can occur during long-term learning of the system.

[0036] In one possible design, the construction of multimodal empirical tuples of visual-physical features and their storage in the evolutionary memory bank in S4 includes: A difference measurement analysis was performed on the code abstract syntax trees of failed and successful iteration steps to identify the core modification actions that led to simulation convergence. The core modification actions, error message scene text, and intermediate visual features are packaged and encapsulated into multimodal experience tuples and stored in the evolutionary memory; the intermediate visual features include mesh distortion or stress concentration images that trigger the error. A dynamic confidence weight is initialized for each multimodal empirical tuple. In subsequent task reuse, if the empirical tuple enables the simulation to converge successfully, the weight is increased; if it causes a secondary error or is not used for multiple consecutive rounds, the weight is decreased. When the weight is lower than a preset threshold, the empirical tuple is automatically pruned and removed.

[0037] S5: Knowledge Graph Topology Scanning and Autonomous Cognition Boundary Exploration.

[0038] In the embodiments of this application, such as Figure 6As shown, the system periodically traverses a multidimensional knowledge graph constructed based on a knowledge base and a memory. A density-based clustering algorithm is used to evaluate the node distribution in the multidimensional feature space: the embedding model is used to calculate the average similarity score of the 10 nearest neighbors to each node in the graph, and the 5 nodes with the lowest similarity scores are marked as knowledge sparse regions. For example, the system detects a lack of verification for the extreme condition of "high fuel consumption accompanied by power transients." The system transforms the data gap features of the sparse regions into physical deduction intentions, automatically synthesizes targeted standard natural language instructions using built-in physical laws, and seamlessly feeds them back to the aforementioned multi-agent framework for execution, achieving extreme condition detection without human intervention.

[0039] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise expressly stated herein, there is no strict order restriction on the execution of these steps, and they may be executed in other orders.

[0040] Specific limitations regarding the 3D modeling and simulation system for nuclear fuel performance analysis can be found in the methodological limitations section above, and will not be repeated here. Each module in the above system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independent of the processor in a computer device, or stored in software in the memory of the computer device, so that the processor can call and execute the corresponding operations of each module.

[0041] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements a transformer fault root cause analysis method based on large-model iterative reasoning. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device casing, or an external keyboard, touchpad, or mouse.

[0042] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0043] In one embodiment, a computer device is provided, which may be a terminal. The computer device includes a processor, memory, a communication interface, a display screen, and an input device connected via a system bus. When executed by the processor, the computer program performs the steps described in the above method embodiments.

[0044] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0045] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A three-dimensional modeling and simulation method for nuclear fuel performance analysis, characterized in that, The method includes: S1. Obtain multimodal heterogeneous documents containing nuclear fuel experimental data. Based on the on-demand triggering strategy and dynamic code probe, perform multimodal knowledge extraction and multi-source consistency verification on the multimodal heterogeneous documents to obtain high-fidelity standard triples. Construct a structured physical parameter knowledge base based on the standard triples. S2, obtain simulation task instructions in natural language form, parse the instructions based on the multi-agent collaborative mechanism and retrieve the corresponding parameters from the physical parameter knowledge base to generate a 3D modeling simulation script; S3, based on the simulation script, set the running parameters of the safety sandbox, collect the physical state results and multimodal intermediate state data output by the solver in real time, if an error occurs, perform error analysis and correction on the error information based on physical rules, and iteratively correct the 3D modeling simulation script until execution is successful; S4 compares the intermediate state data of successful iteration steps and failed iteration steps, extracts the core modification actions that lead to the reversal of the simulation state through difference analysis, encapsulates them into multimodal experience tuples containing visual-physical features, and stores them in the evolutionary memory bank for retrieval and reuse in subsequent modeling tasks. S5. Construct a knowledge graph based on the physical parameter knowledge base and the evolutionary memory base, scan the high-dimensional parameter space of the knowledge graph to identify sparse knowledge regions, and autonomously generate and execute verification tasks under extreme conditions based on the identification results.

2. The method according to claim 1, characterized in that, The multimodal knowledge extraction in S1, based on an on-demand triggering strategy and dynamic code probes, includes: By using a document parsing model, the text and tables in the multimodal heterogeneous documents are decoupled from their layout and restructured to obtain a unified Markdown text format. For the image data in the multimodal heterogeneous document, a global semantic summary of the image data is extracted using a large visual language model. The global semantic summary is then concatenated with the original image caption to construct a cross-modal placeholder tag, which is then embedded in the Markdown text format in situ. Upon receiving the target parameter extraction instruction, a candidate context subset containing the target parameter is obtained from the Markdown text format based on a dual-path retrieval strategy; if the candidate context subset contains the cross-modal placeholder label and the numerical information is missing, the visual language big model is triggered to extract local numerical features from the original high-resolution image cropped block. For unlabeled power history files, the header sequence of the power history file is extracted as metadata prior. A prompt word template is constructed by combining it with the entity name of the target parameter. An abstract syntax tree parsing script is dynamically generated using a large language model. The parsing script is executed in a security sandbox to output the specific value of the target parameter.

3. The method according to claim 2, characterized in that, The dual-path retrieval strategy for obtaining a subset of candidate contexts containing the target parameter in the Markdown text format includes: The BM25 algorithm is used to calculate the keyword matching score between the target parameter extraction instruction and the Markdown text format segmentation block; The instructions and text blocks are transformed into dense vectors using a domain-fine-tuned embedding model, and the cosine similarity score between them is calculated. The keyword matching score and the cosine similarity score are weighted and fused, and the Top-K text blocks are recalled as candidate context subsets based on the ranking result of the fused scores.

4. The method according to claim 1, characterized in that, In step S1, the multi-source consistency verification includes: For multiple candidate observations of the same physical parameter extracted in different modes, their units are uniformly converted to the preset standard dimensions to obtain a normalized numerical sequence. Calculate the variance of the normalized numerical sequence. If the variance is lower than a preset engineering tolerance threshold, then take the mean as the standard value. If the variance is greater than or equal to the engineering tolerance threshold, the softmax average probability of the information source modality token of each observation is calculated as the prior confidence weight, and the average value is taken after removing low confidence noise points with weights lower than the preset value.

5. The method according to claim 1, characterized in that, The multi-agent cooperative mechanism in S2 consists of a planning agent, an encoding agent, an execution agent, and a reflective agent. The planning agent is used to decompose the simulation task instructions into a sequence of subtasks in a directed acyclic graph, and scan the tool operator library to reuse operators or generate new operator development instructions. The coded agent is used to combine the parameters retrieved from the knowledge base with the output of the planning agent to generate the 3D modeling simulation script with type checking and geometric topological constraints. The execution agent is used to call the underlying finite element solver to run the script in an isolated security sandbox, and to monitor the solver's intermediate state logs, convergence residual curves and mesh distortion images in real time, and generate an execution status feedback set. The reflective agent is used to receive the execution state feedback set and perform root cause analysis by combining the built-in nuclear physics prior rules and solver error semantics.

6. The method according to claim 5, characterized in that, The error analysis and correction of the error message based on physical rules in S3 includes: The reflective agent generates a reflective report containing specific suggestions for modifying parameters or adjusting the grid. The reflection report is sent back to the coding agent, which then drives the coding agent to regenerate the 3D modeling simulation script for the next iteration, until the simulation physical state results satisfy the physical conservation law.

7. The method according to claim 1, characterized in that, The construction of multimodal empirical tuples of visual-physical features in S4 and their storage in the evolutionary memory bank includes: A difference measurement analysis was performed on the code abstract syntax trees of failed and successful iteration steps to identify the core modification actions that led to simulation convergence. The core modification actions, error message scene text, and intermediate visual features are packaged and encapsulated into multimodal experience tuples and stored in the evolutionary memory; the intermediate visual features include mesh distortion or stress concentration images that trigger the error. A dynamic confidence weight is initialized for each multimodal empirical tuple. In subsequent task reuse, if the empirical tuple enables the simulation to converge successfully, the weight is increased; if it causes a secondary error or is not used for multiple consecutive rounds, the weight is decreased. When the weight is lower than a preset threshold, the empirical tuple is automatically pruned and removed.

8. A three-dimensional modeling and simulation system for nuclear fuel performance analysis, characterized in that, Performing any one of claims 1-7 as described above includes: The knowledge extraction module is used to acquire multimodal heterogeneous documents containing nuclear fuel experimental data. Based on the on-demand triggering strategy and dynamic code probe, it performs multimodal knowledge extraction and multi-source consistency verification to obtain high-fidelity standard triples and build a knowledge base. The multi-agent modeling module is used to acquire simulation task instructions, parse the instructions based on the multi-agent collaborative mechanism, retrieve parameters from the knowledge base to generate a 3D modeling simulation script, execute it in a safe sandbox, and reflect and iterate based on physical rules when errors occur. The memory management module is used to perform multidimensional distillation on intermediate state data of successful and failed executions, extract core modification actions, encapsulate them into visual-physical multimodal experience tuples and store them in the evolutionary memory bank, and perform dynamic confidence pruning. The autonomous exploration module is used to construct a knowledge graph based on the knowledge base and evolutionary memory, scan and identify sparse knowledge regions, and autonomously generate and execute verification tasks under extreme conditions.

9. A computer device, characterized in that, It includes a memory and a processor, wherein the processor executes a program to implement the method according to any one of claims 1-7.

10. A computer-readable storage medium having a program stored thereon, which, when executed, implements the method of any one of claims 1-7.

Images