RPI Researchers Harness Agentic AI for Smarter, Faster Aerospace Design

May 1, 2026

Rensselaer Polytechnic Institute, Troy, NY

A Rensselaer Polytechnic Institute (RPI) engineering professor, Shaowu Pan, Ph.D. and his team of students have integrated agentic AI into computational fluid dynamics (CFD) to optimize the aerospace design process and alleviate bottlenecks.

Pan’s advances address priorities outlined in Winning the Race: America’s AI Action Plan, which emphasizes that “high-quality data has become a national strategic asset” and calls for “the world’s largest and highest quality AI-ready scientific datasets.”

Supported by funding from Google and the U.S. Department of Energy, Pan’s team achieved three major advances in 2025: they created Unifoil, a massive airfoil simulation dataset, developed a large language model (LLM) framework capable of running CFD simulations, and built benchmark tools to evaluate LLM accuracy on CFD tasks.

Unifoil – World’s Largest Airfoil Simulation Dataset

In collaboration with a research group under Professor Sicheng He from The University of Tennessee, Knoxville, Pan’s team developed UniFoil, the world’s largest RANS-based airfoil simulation dataset. The dataset contains more than 500,000 samples across diverse flight conditions.

“A huge dataset with challenging aerospace problems is desperately needed so that people from the AI community can use this data to evaluate their models that were previously designed for processing images and videos,” Pan said. “This would speed up the adoption of AI models in the aerospace community and eventually lead to smarter and faster aerospace designs.”

The UniFoil dataset establishes a benchmark for training and evaluating aerodynamic models. “The UniFoil dataset is challenging because it includes laminar–turbulent transition effects, as well as compressible transonic regimes where shock waves form on the airfoil surface, leading to strong pressure gradients and shock–boundary-layer interactions. These phenomena introduce interesting nonlinearities in the function space that can be difficult to capture,” explains RPI aerospace engineering Ph.D. student Nithin Somasekharan, a member of Pan’s team.

Foam-Agent – A Multi-Agent LLM System Automating CFD Workflows

To reduce the labor involved in automating CFD workflows, Pan’s RPI team also created Foam-Agent, a multi-agent LLM system that automates computational fluid dynamics workflows from natural language instructions. “You basically bring ChatGPT intelligence into the design phase of the production cycle,” said Ling Yue, a computer science Ph.D. student and another member of Pan’s team. “The framework automates complex simulations, democratizing scientific computing by lowering the expertise barrier for computational fluid dynamics.”

CFDLLMBench - First Benchmark Suite to Evaluate LLMs

To provide feedback on better model design, Pan’s research group, along with researchers Patrick Emami, Ph.D., of National Laboratory of the Rockies, and Anurag Acharya, Ph.D., of Pacific Northwest National Laboratory, further developed the first comprehensive benchmark suite for evaluating LLMs on computational fluid dynamics tasks. Named CFDLLMBench, the benchmark holistically evaluates whether an LLM knows graduate-level CFD concepts, can do numerical/physical reasoning, and whether it can implement context-dependent CFD workflows.

Pan and his colleagues hope the three advances transform how engineers approach computational fluid dynamics, a notoriously complex field with a high barrier for entry.

Aerospace America, a trade journal published by the American Institute of Aeronautics and Astronautics (AIAA), recently recognized this body of work among 2025’s most significant aerospace advances in its annual “Year in Review.” This recognition reflects the strength of RPI’s aerospace expertise and demonstrates how agentic AI is transforming engineering workflows, even in some of the most challenging disciplines.

Pan, who brought his unique background in aerospace engineering, computer science, applied mathematics and scientific machine learning to RPI in 2022, is determined to tackle some of the most frustrating problems in aerospace. “I had the fortune to have explored a wide range of the hardest problems in aerospace over the past 15 years such as hypersonic nonequilibrium flows, turbulent combustion, and compressible turbulence.” Pan said. “If I want to find disruptive ideas to transform this field, I spend time on mathematics and AI.”

Pan’s work represents what he calls “a philosophical pivot” in how we tackle engineering’s toughest puzzles. “For the last few decades, we computational scientists have basically been acting as pro-bono consultants for our own hardware,” he said. “We spent decades burning our own brain cells to write ‘clever’ algorithms just so our computers wouldn’t have to work a sweat. We were doing the heavy lifting to keep the CPUs comfortable. Now, we’re finally letting the computers take the wheel on the grunt work, freeing us up to handle the finishing touches, the high-level creativity that actually requires a brilliant mind.”

This article was written by Chris Ingraham for the Rensselaer Polytechnic Institute. For more information, contact Chris, This email address is being protected from spambots. You need JavaScript enabled to view it..