Differentiable physics-enabled closure modeling for Burgers’ turbulence

Abstract Data-driven turbulence modeling is experiencing a surge in interest following algorithmic and hardware developments the data sciences. We discuss an approach using differentiable physics paradigm that combines known with machine learning to develop closure models for Burgers’ turbulence. consider one-dimensional Burgers system as prototypical test problem unresolved terms advection-dominated problems. train series of incorporate varying degrees physical assumptions on posteriori loss function efficacy across range parameters, including viscosity, time, grid resolution. find constraining inductive biases form partial differential equations contain or existing approaches produces highly data-efficient, accurate, generalizable models, outperforming state-of-the-art baselines. Addition structure information also brings level interpretability potentially offering stepping stone future modeling.