Announcement Detail

Friday, May 1, 2026

10:00 AM CDT

Join via Zoom: https://us06web.zoom.us/j/83559598613?pwd=nyFbCjIt4cl0TBjT558DLRaNRfxMz1.1

CFD/FSI Webinar Series

Differentiable Hybrid Neural-PDE Operator for Modeling of Spatiotemporal Physics

Prof. Jian-Xun Wang, Cornell University

Abstract:

This talk presents neural differentiable modeling for fluid dynamics: a unified framework that embeds governing PDE operators as differentiable, GPU-accelerated layers within trainable computational graphs, enabling end-to-end coupling of high-fidelity CFD solvers with machine learning. The framework preserves conservation laws, numerical stability, and physical structure, while enabling gradient-based learning of closures, boundary conditions, and latent operators.

A central contribution is the development of a differentiable CFD solver stack for complex flows, spanning both structured and unstructured discretizations. On structured grids, we introduce Diff-FlowFSI, a fully vectorized, GPU-native differentiable CFD/FSI platform built upon a sharp-interface immersed boundary method (IBM), enabling accurate simulation of complex and moving geometries with end-to-end differentiability through boundary treatments, fluxes, and time integration. On unstructured meshes, we develop DiFVM, which reformulates finite-volume operators as graph-based message-passing primitives, enabling scalable, end-to-end differentiable simulation on arbitrary geometries with OpenFOAM-compatible workflows. Together, these solvers establish a unified differentiable infrastructure for high-fidelity flow simulation.

Building on this foundation, we present hybrid neural–CFD models for turbulent flows, where physically constrained learning of subgrid-scale closures and wall models enables robust zero-/few-shot generalization across Reynolds numbers, geometries, and forcing conditions. Generative diffusion models are further integrated with differentiable solvers to perform super-resolution and posterior sampling of spatiotemporal flow fields from sparse observations. Beyond fluid mechanics, the framework extends to multiphysics and multiscale transport, including phonon Boltzmann transport (BTE) and coupled BTE–Fourier models for chip-scale thermal digital twins, where differentiable solvers enable cross-scale coupling, inverse modeling, and real-time thermal prediction. Additional applications include manufacturing processes such as chemical vapor infiltration (CVI) and electrochemical machining (ECM), where differentiable transport–reaction and moving-boundary models enable design-in-the-loop optimization and uncertainty-aware control. Across all applications, calibrated uncertainty quantification is tightly integrated within the differentiable pipeline.

Bio:

Dr. Jian-Xun Wang is an Associate Professor in the Sibley School of Mechanical and Aerospace Engineering at Cornell University. He received his Ph.D. in Aerospace Engineering from Virginia Tech in 2017 and completed postdoctoral training at UC Berkeley before joining the University of Notre Dame as a tenure-track Assistant Professor in 2018. He is a recipient of the NSF CAREER Award and the ONR Young Investigator Program (YIP) Award, and currently serves as an Associate Editor for the Journal of Computational Physics and as Vice Chair of the USACM Technical Thrust Area on Data-Driven Modeling.

Dr. Wang’s research lies at the intersection of scientific machine learning, computational fluid, solid, and thermal dynamics, data assimilation, and uncertainty quantification. His work aims to advance predictive modeling, inverse analysis, and decision-making for complex physical systems.