关键词: 多智能体系统, 科学人工智能, 核物理, 多模态知识抽取, 智能体记忆

Abstract: The rapid advancement of Large Language Models (LLMs) has catalyzed the paradigm of Artificial Intelligence (AI) for Science in multi-physics simulations. However, in the rigorous domain of nuclear fuel performance analysis, high-fidelity 3D thermo-mechanical modeling still heavily relies on manual expertise for heterogeneous data extraction and operator development. Existing multi-agent systems face three critical bottlenecks: multimodal “hallucinations” when parsing long-tail documents, repetitive trial-and-error due to the lack of physical constraints, and passive dependence on human instructions that blinds them to hidden physical boundaries. To bridge this gap, a closed-loop Multi-Agent 3D Modeling and Simulation Validation System for nuclear fuel performance analysis was proposed. First, we design an on-demand multimodal knowledge extraction pipeline equipped with dynamic code probes and multi-source consistency verification to extract high-fidelity physical parameters from unstructured archives. Second, we develop an auto-expanding multi-agent framework featuring a multimodal evolutionary memory mechanism. Operating within a secure sandbox, it forms a "generation-reflection-encapsulation" loop and innovatively distills "visual-physical" intermediate trajectories (including mesh distortion and other geometric features) into system memory. Coupled with dynamic confidence pruning, this effectively mitigates catastrophic forgetting. Finally, we introduce an autonomous cognitive evolution mechanism that scans knowledge graph topologies to detect sparse regions in high-dimensional parameter spaces, enabling zero-human-intervention exploration of extreme conditions. Evaluated on a real-world nuclear engineering dataset and challenging simulation benchmarks, our system achieves an F1-Score of 93.6% in core parameter extraction. After 50 rounds of evolution, modeling time is drastically reduced to 17.9 minutes. This work not only shatters multimodal data silos in complex engineering but also provides a promising technical path for proactive AI-synergistic simulations in exploring high-dimensional hidden physical domains.

Key words: multi-agent systems, Artificial Intelligence (AI) for science, nuclear physics, multimodal knowledge extraction, agent memory