Research map / Large language models / Agents, code and math
AutoFEA: Enhancing AI Copilot by Integrating Finite Element Analysis Using Large Language Models with Graph Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds AutoFEA, a system that uses step-by-step planning, GCN-Transformer link-prediction retrieval, and LLM code generation to produce CalculiX FEA input files, validated on 512 projects.
Problem
LLMs hallucinate when generating complex, specialized engineering simulation inputs such as FEA files, producing plausible but physically incorrect results for non-expert users.
Approach
Decompose FEA code generation into planned steps, retrieve relevant example code via a GCN-Transformer link predictor over a step graph (edges linking steps sharing keywords), and ground the LLM's outputs by running the generated CalculiX simulation.
Key results
- Across five settings, the full AutoFEA pipeline achieves the highest simulation success rate, outperforming direct generation, RAG, planning-only, and retrieval-augmented planning baselines (Figure 3).
- With GPT-4o, AutoFEA reaches a 90.2% success rate; GPT-4o-mini 81.6%, Gemini-1.5-Pro 83.8%, Gemini-Flash 75.2%, Llama 3 8B 78.8%.
- In a cantilever beam case study, GPT-4o's linear-analysis answer (~0.0975) was corrected to ~0.8738 after feeding back a CalculiX simulation that included plasticity.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AutoFEA's planning, retrieval, code-generation, and simulation pipeline | LLM-based FEA code generation · FEA projects test set | simulation success rate | Helped |
| replace GPT-4o with GPT-4o-mini (ablation) | AutoFEA · FEA projects test set | success rate (%): 90.2 -> 81.6 | Hurt |
| replace GPT-4o with Gemini-Flash (ablation) | AutoFEA · FEA projects test set | success rate (%): 90.2 -> 75.2 | Hurt |
| replace GPT-4o with Gemini-1.5-Pro (ablation) | AutoFEA · FEA projects test set | success rate (%): 90.2 -> 83.8 | Hurt |
| replace GPT-4o with Llama 3 8b (ablation) | AutoFEA · FEA projects test set | success rate (%): 90.2 -> 78.8 | Hurt |
| integrate FEA simulation results into the LLM reasoning process | GPT-4o · Cantilever beam case study | displacement: 0.0975 -> 0.8738 | Helped |
Limitations
- Excluded projects with long codes to limit GPT-4o cost, so the dataset covers only shorter simulations.
- Evaluation relies on a single CalculiX test-example source and a success criterion comparing output files; only one hallucination case study.
- Some structure and mesh blocks were truncated beyond 10 lines, which may affect fidelity.
Experimental setup
Datasets: CalculiX official test examples (512 projects)
Benchmarks: AutoFEA 512-project FEA dataset
Metrics: simulation success rate (dat-file comparison with relative t
Baselines: Direct Generation Retrieval-Augmented Generation Planning-Based Generation Retrieval-Augmented Planning
Builds on: ReAct PAL GNN-RAG
Closest papers by meaning
- AMBER: Adaptive Mesh Generation by Iterative Mesh Resolution Prediction (NeurIPS 2025 · similarity 0.83)
- G-Adaptivity: optimised graph-based mesh relocation for finite element methods (ICML 2025 · similarity 0.83)
- AutoHood3D: A Multi‑Modal Benchmark for Automotive Hood Design and Fluid–Structure Interaction (NeurIPS 2025 · similarity 0.82)
- Diffusion-Based Hierarchical Graph Neural Networks for Simulating Nonlinear Solid Mechanics (NeurIPS 2025 · similarity 0.82)
- CAD-Llama: Leveraging Large Language Models for Computer-Aided Design Parametric 3D Model Generation (CVPR 2025 · similarity 0.82)
- AeroGTO: An Efficient Graph-Transformer Operator for Learning Large-Scale Aerodynamics of 3D Vehicle Geometries (AAAI 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.