Distilling experience into a physically interpretable recommender system for computational model selection

Abstract

Model selection is a chronic issue in computational science. The conventional approach relies heavily on human experience. However, gaining experience takes years and is severely inefficient. To address this issue, we distill human experience into a recommender system. A trained recommender system tells whether a computational model does well or poorly in handling a physical process. It also tells if a physical process is important for a quantity of interest. By accumulating this knowledge, the system is able to make recommendations about computational models. We showcase the power of the system by considering Reynolds-averaged-Navier–Stokes (RANS) model selection in the field of computational fluid dynamics (CFD). Since turbulence is stochastic, there is no universal RANS model, and RANS model selection has always been an issue. A working model recommending system saves fluid engineers years and allows junior CFD practitioners to make sensible model choices like senior ones.