Speaker
Description
Buoyancy-driven turbulence arises across a wide range of natural and technological systems, from inertial confinement fusion to stellar explosions. Direct numerical simulation of all dynamically active scales is often prohibitively expensive, motivating large-eddy simulation and the development of accurate subgrid closures. We extend recent work on a solver-in-the-loop framework in which the governing equations are embedded directly into the training of neural-network closures for unresolved terms. The approach is investigated for homogeneous Rayleigh–Bénard convection and Rayleigh–Taylor turbulence in both two and three dimensions. We evaluate the learned models a posteriori through long-time integrations, focusing on the probability distributions of energy and scalar-variance fluxes, high-order statistics, and the limitations of conventional closure models in reproducing these quantities. Finally, we examine how the unrolled training horizon influences accuracy and stability, relating the required time in the loop to relevant time scales of the flow.