Reinventing Simulation with Intelligent FEA Agents for Bias-Aware Guidance, Traceability, Error Control, Task Automation, and Beyond

We propose an AI agent toolkit to automate finite element analysis (FEA) workflows, reducing reliance on error-prone and time-consuming manual processes. Current simulation methods demand expert input throughout the entire workflow. Bias or errors can be injected from decisions like geometry simplification, mesh resolution, boundary conditions, material models, or other critical assumptions. Our solution will ingest geometry and natural language inputs to generate meshes, guide simulation setup, monitor execution, and post-process results while transparently highlighting trade-offs in critical decisions (e.g., mesh refinement, boundary conditions, element types, material assumptions, etc.).In Phase I, we will:ÝAnalyze error sources in existing workflows, leveraging our extensive experience creating and running finite element simulations.ÝIdentify AI/ML techniques (e.g., reinforcement learning for parameter optimization) to mitigate these errors.Develop a roadmap for Phase II prototype development.Prove the technical feasibility of the agentís core components. Optimized meshing and solver inputs will be generated and executed from natural language prompts.The agent will be a fine-tuned model leveraging cutting-edge large language models (LLMs) with reasoning such as Anthropicís Claude 3.7 Sonnet and OpenAIís o4-mini. This system will automate solid mechanics simulations in Sandia National Laboratoriesí Sierra/SolidMechanics, leveraging Sierraís validation data and meshing tools (Krino, Cubit/Sculpt). The agentís ability to set up a basic problem, execute the simulation, and detect instabilities and insufficient resolution will be demonstrated on exemplar problems in Phase I.Phase I Option and Phase II will mature this prototype AI agent by implementing adaptive decision-making, addressing error sources per the Phase I roadmap, and enabling continuous improvement through user feedback. The final system will reduce manual effort while enhancing accuracy for defense and commercial applications.Ý