CaDS Seminar 2026 - Feb. 17

Tom Jose and Prakhar Rathore (SDL Fluids & Solids Engineering)

Bridging LES and DNS through AI-driven Sub-Grid Scale Modeling

Abstract:

Direct numerical simulations (DNSs) and large-eddy simulations (LESs) provide deep insights into turbulent flows, but at prohibitively high computational costs. While LESs offers a more affordable alternative to DNSs, they inherently lack the sub-grid-scale (SGS) details necessary to precisely describe turbulence. To bridge this gap, our talk presents an AI-driven framework that leverages Super-Resolution (SR) networks to reconstruct high-fidelity or DNS-like fields from low-fidelity datasets generated by LESs. The Simulation and Data Lab - Fluids & Solids Engineering (SD FSE), in collaboration with the University of Magdeburg, is developing such a framework for reconstructing high-fidelity and subsequently SGS turbulence. For this goal, a computational fluid dynamics (CFD) dataset of a human larynx is used, which exhibits complex, transitional patterns during exhalation. By adapting SR techniques from computer vision - specifically a convolutional defiltering model (CDM) utilizing a U-Net architecture - we exploit high-fidelity data to learn representations of unresolved physics directly. Beyond biomedical flows, the SDL FSE team, in collaboration with domain experts from Georgia Tech, Imperial College London, and the University of Zaragoza, has identified and benchmarked open-source, GPU-enabled solvers for the compressible Navier-Stokes equations, where the proposed SR framework will be extended, enabling DNS-like diagnostics at a fraction of the computational cost.