Physics-informed machine learning approach for sub-grid scale modelling in LES

Industrial multiphase flows typically involve turbulent structures on either side of a fluid interface. Modelling efforts are computationally expensive, as large-eddy simulations (LES) require non-trivial near-interface treatments. In this study, we use a physics-informed machine learning (ML) approach for reconstructing LES sub-grid viscosity models. The ML model, trained on a DNS database of low Reynolds number wavy stratified flows, is then evaluated in out-of-sample Reynolds number settings. The ML algorithm can advantageously capture generic small-scale features from computationally expensive flows and can be coupled within each iteration of the physics based flow solver to reduce computational time. We quantify the saving in computational time versus accuracy vis-à-vis traditional dynamic LES models. The study demonstrates how coupling transparent physics based models with data-driven black box machine learning models leverages the best of both methodologies and gives computationally tractable and reliable solutions.