Towards intelligent CFD workflow in the era of large-language models
Large language models are beginning to act as interfaces to scientific software, but in CFD the bar is higher than fluent and sounding text. I will first introduce CFDLLMBench, a comprehensive benchmark for evaluating LLMs on CFD knowledge, numerical reasoning, and context-dependent OpenFOAM workflows. Using this lens, I will then present Foam-Agent, a multi-agent pipeline that couples retrieval, dependency-aware case generation, and iterative error repair; on 110 tasks it reaches 83.6% executable success. Finally, I will discuss FoamGPT, which consists of an open source LLM fine-tuned on a set of curated OpenFOAM tutorials to improve the naive baseline (i.e., generic LLM alone). We close with discussions on failure modes and research opportunities for trustworthy “NL â CFD” automation.
Bio: Shaowu Pan is an assistant professor in the Department of Mechanical, Aerospace, and Nuclear Engineering at Rensselaer Polytechnic Institute (RPI) and leads the Computational Scientific Machine Learning Lab. He is also affiliated with the Rensselaer-IBM Artificial Intelligence Research Collaboration. Pan received his B.E. in aerospace engineering and B.S. in applied mathematics from Beihang University, and his M.S. and Ph.D. in aerospace engineering and scientific computing from the University of Michigan, Ann Arbor in 2021. Prior to joining RPI in 2022, he was a postdoctoral scholar at the AI Institute in Dynamic Systems at the University of Washington. His research focuses on scientific machine learning, operator-theoretic modeling and control of nonlinear systems, and data-driven modeling of large-scale physical systems, with applications in computational fluid dynamics and clean energy. He is a recipient of the 2025 Google Research Scholar Award.