Speaker
Description
The evolve-filter (EF) strategy is a spatial filter-based numerical stabilization technique for convection-dominated flows. In under-resolved regimes, classical numerical approaches may lack accuracy due to the presence of spurious numerical oscillations. EF offers a simple, modular, and effective approach to smooth out these instabilities.
However, it is well known that when the filter action is too strong, the method may lead to inaccurate and over-diffusive results.
In this talk, we present novel strategies to mitigate the over-diffusivity of EF when large filter radii are employed, while preserving the main flow features. We explore two complementary directions: modifying the model and optimizing its parameters.
From an algorithmic perspective, we propose a new approach based on the variational multiscale (VMS) framework, which allows us to separate the resolved large scales from the resolved small scales in the evolved velocity field. The filtered small scales are then used to correct the large scales, enhancing accuracy without losing important flow information.
From an optimization perspective, we demonstrate the crucial role of the filter parameter and introduce a reinforcement learning strategy for its selection. Notably, the training of the network does not rely on direct numerical simulations, significantly reducing the computational cost of the learning procedure.