On statistical solutions of fluids and their computation (Online)

Abstract

We start by demonstrating that numerical methods do not necessarily converge to entropy or admissible weak solutions of the Euler and Navier-Stokes equations of fluid dynamics on mesh refinement due to appearance of eddies at smaller and smaller scales. As an alternative, we revisit the concept of statistical solutions which are time-parametrized probability measures, consistent with the fluid evolution. We empirically show that the same numerical methods converge to a statistical solution and also derive verifiable sufficient conditions under which this convergence can be made rigorous. Numerical experiments illustrating interesting properties of statistical solutions are also presented. We conclude by showing how state of the art generative AI models (conditional diffusion) can significantly lower the cost of computing statistical solutions while maintaining accuracy.