SISSA Trieste - Italy
The talk will focus on a novel data-driven surrogate modeling approach for defining the filter and relax steps within the Evolve–Filter–Relax (EFR) framework for turbulent flow simulations. Standard EFR methods rely on a small number of tunable parameters, which limits their flexibility and adaptability to different flow regimes. In contrast, our approach learns an optimal filtering operator directly from DNS data in the frequency domain, resulting in a flow-informed surrogate filter. The learning procedure is computationally efficient, as it involves only a set of independent one-dimensional least-squares problems, one for each wavenumber. We demonstrate the effectiveness of the proposed method on decaying turbulence and Kolmogorov flow, where it significantly improves the accuracy of energy spectra and the time evolution of energy and enstrophy compared to classical filtering techniques. In addition, the relax parameter is selected by enforcing energy and/or enstrophy conservation, which enhances stability and reduces numerical oscillations, particularly in data-scarce settings. Finally, the learned surrogate filter is also computationally cheaper to apply than traditional differential filters, as it avoids the solution of a linear system.