Hybrid Physics-Machine Learning Framework Toward Efficient Simulation of Turbulent Flows on an Exascale Platform

We present a hybrid physics–machine learning (ML) framework for predicting time-dependent fluid flow. This framework directly integrates ML into existing PDE-based simulations, significantly accelerating computations compared to high-fidelity simulations. By approximating small-scale flow dynamics, our method preserves accuracy while reducing computational costs. Our approach employs convolutional neural networks (CNNs), with a loss function derived from projected solutions of high-resolution simulations. Training is conducted within the simulation environment. We validate our CNN approach using backward-facing step flows using the GPU-accelerated NekRS solver on the recently deployed exascale platform, Aurora, at Argonne National Laboratory. Our results highlight the potential of combining physics-based modeling and machine learning to enhance computational efficiency and accuracy in turbulence research.