VFEAgent: AI Agents Automate Finite Element Analysis from Images

Summary

Finite Element Analysis (FEA) is a cornerstone of modern engineering design, but its workflow remains notoriously complex and expertise-intensive. While Large Language Models (LLMs) have been applied to FEA, they struggle with multimodal inputs (e.g., images of geometries) and complex multi-step tasks. VFEAgent addresses these limitations by introducing an end-to-end multi-agent system that automates FEA modeling and simulation directly from input images and problem descriptions.

The system integrates two core components: a multimodal vision-language multi-agent pipeline that uses ReAct-driven reasoning to extract structured FEA specifications from heterogeneous inputs, and a verification-first code synthesis framework with self-debugging and fallback mechanisms to ensure executability and physical validity. Evaluated across various engineering mechanics scenarios, VFEAgent achieves high success rates in generating complete and physically valid simulations, significantly outperforming LLM-based baseline methods in both reliability and correctness.

Key Contributions

• End-to-end automation of FEA: First system to automate the complete FEA workflow from multimodal inputs (images + text) to executable simulation code.
• Multimodal vision-language agent pipeline: Novel use of ReAct-driven reasoning to extract structured FEA specifications from heterogeneous inputs, enabling the system to understand geometry from images.
• Verification-first code synthesis: A robust framework incorporating self-debugging and fallback mechanisms that ensures generated simulation code is both executable and physically valid.
• Empirical validation: Systematic evaluation across multiple engineering mechanics scenarios demonstrating superior performance over LLM-based baselines.

Implications

For Researchers

This work provides a validated architecture for automating complex engineering simulations, opening new research directions in multimodal agent systems for scientific computing. The verification-first approach offers a template for ensuring reliability in agent-generated code for high-stakes domains. Researchers can build on this framework to extend automation to other physics-based simulations (CFD, electromagnetics) and explore more sophisticated multi-agent coordination patterns.

For Developers

VFEAgent demonstrates a practical pattern for building domain-specific multi-agent systems that combine vision, reasoning, and code generation with robust error handling. The architecture—using specialized agents for different workflow stages with a verification-first design—is directly applicable to other engineering automation tasks. Developers can adopt the self-debugging and fallback mechanisms to improve reliability in code-generation agents.

For Users

Engineers and designers can potentially automate tedious FEA setup and analysis, reducing manual effort and enabling faster design iteration cycles. The system lowers the barrier to entry for FEA, allowing users without deep domain expertise to generate valid simulations. This could democratize access to simulation-driven design across smaller engineering firms and educational settings.

References

• https://arxiv.org/abs/2605.28978v1