Agentic AI
New Operating Model Lets AI Coding Agents Manage CFD Simulations with Human Review
Researchers demonstrate a structured framework for deploying large language models to develop and automate computational fluid dynamics workflows in OpenFOAM, with built-in engineering oversight and reusability.
Coding agents powered by large language models can now systematically develop and execute computational fluid dynamics (CFD) simulations while maintaining human control and review. A new case-bundle operating model, demonstrated in an OpenFOAM-based study, separates the agent-assisted design phase from automated execution, enabling both engineering judgment and scalable reuse.
The Two-Mode Framework#
The approach uses a dual-process model: Build mode and Replay mode. In Build, LLM-based coding agents assist engineers in developing a complete case bundle—essentially a self-contained package of simulation configuration, geometry processing, meshing procedures, execution scripts, post-processing logic, and documentation. Once reviewed and approved, the bundle enters Replay mode, where the same setup can be applied to new problem variants automatically, without requiring re-engineering.
Injector Design Study#
Researchers applied this framework to screen injector designs using OpenFOAM's multiphase flow solver (interFoam). GPT-5.5 in Codex mode helped assemble the case bundle during development. The reviewed bundle was then replayed to run and analyze 140 different geometry variants defined by Stereolithography (STL) files on a remote supercomputing system. This demonstrated how a single vetted workflow can scale to handle substantial parameter sweeps.
Multi-Backend Validation#
A second replay exercise tested the same case bundle with Pi coding agent as the runtime, paired with four different large language model backends. All four configurations executed successfully and produced verified results, though tool usage and token consumption differed. This flexibility shows the framework can work across different AI infrastructure choices while maintaining workflow consistency.
Why It Matters#
The model addresses a practical gap in AI-assisted scientific computing: coding agents can automate tedious simulation setup, but results need engineering oversight and reproducibility. By treating case development and execution as separate concerns—with human review acting as a gate between them—the framework enables bounded, trustworthy automation that respects both software engineering rigor and domain expertise.
