Minimal error partially resolving simulation methods for turbulent flows: A dynamic machine learning approach

Abstract

A significant extension of previously introduced continuous eddy simulation methods is presented by introducing minimal error partially and fully resolving simulation methods for turbulent flows. This approach represents a machine learning strategy for the hybridization of modeling-focused and resolution-focused simulation methods. It can be applied to well-known equation structures (Spalart–Allmaras type equations, usually applied two-equation models), and it can be used for different hybridization types and in different computational versions. Physically, minimal error methods implement a mode interplay, which ensures that the resolution imposed by a model equals the actual flow resolution. Differently formulated simulation methods reveal two typical errors, and they cannot be expected to provide reliable predictions under conditions where validation data are unavailable. These problems can be avoided by minimal error formulations of model structures considered.