Subgrid closure for the Sabra shell-model cascade, from scratch in JAX. A one-parameter eddy viscosity beats a learned MLP that is twice as good offline - what sets the spectrum is the mean energy drain, not the pointwise fit.
deep-learning computational-physics turbulence computational-fluid-dynamics jax shell-model large-eddy-simulation scientific-machine-learning closure-modeling subgrid-scale-model
-
Updated
Aug 5, 2026 - Python