Towards a deep learned subgrid-scale surrogate model for stratified turbulence from high-resolution simulation data

Abstract:

The abundance of multi-scale physics constituting turbulent flows prohibitively increases the computational cost for simulating such flows, particularly with the added complexity of density stratification. Thus, modeling complex atmospheric and climate phenomena involving cloud micro-physics and turbulence has always been a challenge. This has been dealt with by using cloud resolving models (CRM) which involves a high-resolution cloud model, parametrizing deep convection, being embedded into each grid cell of a coarse resolution global climate model. This approach is known as Multi-scale Modeling Framework (MMF), utilized in the U.S. DOE Energy Exascale Earth System - Model Multi-scale Modeling Framework (E3SM-MMF) for global climate modeling. On a similar note, large eddy simulations (LES) rely on filtering the high-wavenumber (fine-scale) structures and resolving only a coarser domain with embedded subgrid-scale (SGS) modeling to parametrize the viscous dissipation from the filtered scales. Although, this is limited by the appropriate choice of the parametrization, mainly consisting of an eddy-viscocity kernel.

Recently, supervised machine learning techniques, relying on components in the eddy-viscocity term of the filtered Navier-Stokes equation, have been introduced to parametrize the SGS model from high-resolution data for LES. Also, deep learning techniques have been used to parametrize the subgrid-scale processes in climate models, which concentrate on capturing the dissipative effects. We use tools from deep learning to generate a SGS surrogate model for stratified turbulence to represent and predict fine scale physics on coarse grid simulations. Starting from the Euler equations with density stratification, we generate subgrid-scale data as a difference between high- and low-resolution flow field data. The dissipative and anti-dissipative components are embedded into the coarse simulation as it is integrated with the high-resolution simulation as a parallel driver. The deep neural network model is generated with the high-resolution grid data as the input and the SGS data as the output. The present supervised approach serves as a novel alternative to the SGS models of CRM, encapsulating both dissipative and anti-dissipative effects of the subgrid-scale.