Title: Arbitrary-Geometry Neural Operators for AI-Driven Design of Energy Flow Components
Abstract: Designing energy-flow components, such as battery electrodes, wind turbines, and heat-transfer devices, often requires expensive high-fidelity simulations to understand how fluids and other physical fields respond to changes in geometry. Machine learning surrogates offer the potential to accelerate these predictions. Most existing surrogate models, however, are limited to a fixed family of geometries or meshes, restricting the range of designs they can efficiently explore. This talk presents a class of geometry-aware neural operators that directly incorporate geometry as an input. We introduce ArGEnT, a transformer-based framework developed through an interlaboratory collaboration among PNNL, LLNL, and SNL. ArGEnT learns a single model capable of predicting physical fields across diverse geometries and discretizations. We demonstrate its performance on fluid-flow and solid-mechanics problems, including turbulent airfoil flows, where the model generalizes to previously unseen geometries and provides near-real-time predictions. We further discuss how geometry-aware neural operators can enable rapid design exploration and optimization over free-form shapes, and we highlight open challenges in model reliability, uncertainty quantification, scalability, and the theoretical understanding of neural operators for complex engineering systems.
Bio: Yucheng Fu is a mathematician at Pacific Northwest National Laboratory. He received his PhD in mechanical engineering from Virginia Tech and his BS in theoretical and applied mechanics from Fudan University. His research focuses on multiphysics modeling, computational energy storage, and scientific machine learning. He has published more than 40 peer-reviewed articles and serves on the Editorial Board of Scientific Reports. He currently leads the DOE Genesis Phase I project GOLD-FLOW, developing AI-enabled methods for fluid-flow design and optimization, and also leads interlaboratory research on physics-informed neural operators in collaboration with LLNL and SNL.