Physics-informed data-driven inference of an interpretable equivariant LES model of two-dimensional incompressible fluid turbulence

Restrictive phenomenological assumptions limit the accuracy and generality of subgrid-scale models of fluid turbulence, especially in flows where coherent structures and bidirectional energy transfer play an important dynamical role. This paper introduces a symbolic, physically-informed data-driven subgrid-scale closure for two-dimensional incompressible turbulence. The closure is inferred from filtered direct numerical simulation data using symmetry-constrained sparse regression, together with dimensional-consistency constraints on the inferred coefficients. While the model is inferred using LES-style spatial coarse-graining, its structure is found to be more similar to RANS models, as it employs an additional field to describe the dynamics of unresolved scales. For the range of two-dimensional benchmark flows considered here, the inferred model substantially improves upon widely used LES models, yielding accurate predictions of subgrid-scale stresses, local and global fluxes between different scales, and other key quantities in both a priori and a posteriori benchmarks.