Autonomous Simulation: ALL-FEM Revolutionizes Finite Element Workflows

Finite element analysis (FEA) has long been the backbone of computational engineering, central to the design and verification process of manufactured objects. Yet, the domain remains complex, requiring deep expertise in numerical analysis and programming. Conventional large language models (LLMs) have been employed to generate FE code, but they often falter, lacking the necessary awareness of variational structures.

Introducing ALL-FEM

Enter ALL-FEM, a novel autonomous simulation system that merges agentic AI with domain-specific, fine-tuned LLMs. This system is designed for FEniCS code generation, applicable across solid, fluid, and multiphysics applications. By building a corpus of over 1000 verified FEniCS scripts, ALL-FEM fine-tunes LLMs with parameters ranging from 3 billion to 120 billion. It combines 500 curated expert codes with a retrieval-augmented, multi-LLM pipeline to generate and filter codes for various PDEs, geometries, and boundary conditions. The precision and depth of this approach mark a significant leap in the field of computational science.

Performance and Benchmarks

ALL-FEM was evaluated on 39 benchmarks, addressing challenges such as linear and nonlinear elasticity, Newtonian and non-Newtonian flow, and fluid-structure interaction. The standout performer, the fine-tuned GPT OSS 120B model, achieved a code-level success rate of 71.79%, surpassing the non-agentic deployment of GPT 5 Thinking. The key contribution here's the demonstration that relatively small, fine-tuned LLMs can efficiently automate complex finite element workflows.

Significance and Future Implications

But why should this matter to the engineering community? The crux of ALL-FEM’s innovation lies in its potential to democratize access to sophisticated simulation tools. By lowering the barrier to entry, it empowers engineers to focus on design and innovation rather than grappling with the intricacies of coding. Can this be the catalyst for a broader transformation in computational engineering?

Critically, ALL-FEM not only automates but also enhances the reliability and accuracy of simulations through its multi-agent workflow and embedded runtime feedback. This builds on prior work from various computational science advancements, paving the way for future applications in other fields reliant on simulation technology.

ALL-FEM offers a blueprint for autonomous simulation systems, setting a new standard in efficiency and capability. The broader implications for engineering and beyond are vast, as the integration of AI into simulation processes becomes increasingly indispensable.