LAUSR

Abstract This work presents developments to a novel evolutionary framework that symbolically regresses algebraic forms of the Reynolds stress anisotropy tensor. This work contributes to the growing trend in machine-learning for modelling physical phenomena. Our framework is shown to be computational inexpensive and produce accurate and robust models that are tangible mathematical expressions. This transparency in the result allows us to diagnose issues with the regressed formulae and appropriately make amendments, as we further understand the regression tools. Such models are created using hybrid RANS/LES flow field data and a passive solving of the RANS transport equations to obtain the modelled time scale. This process shows that models can be regressed from a qualitatively correct flow field and fully resolved DNS is not necessarily required. Models are trained and tested using rectangular ducts, an example flow genus that linear RANS models even qualitatively fail to predict correctly. A priori and a posteriori testing of the new models show that the framework is a viable methodology for RANS closure development. This a posteriori agenda includes testing on an asymmetric diffuser, for which the new models vastly outperform the baseline linear model. Therefore this study presents one of the most rigorous and complete CFD validation of machine learnt turbulent stress models to date.