LES informed data-driven models for RANS simulations of single-hole cooling flows

A LES-informed data-driven approach for improved predictions of the turbulent heat flux vector has been sought for film and effusion cooling flow applications. Random forest and shallow neural networks have been used to train a spatially varying coefficient for the Higher-Order Generalised Gradient Diffusion Hypothesis (HOGGDH) turbulent heat flux closure model. a priori results of the turbulent heat flux magnitude showed significant improvements over the standard HOGGDH model. The random forest model was implemented into OpenFOAM with a previously published data-driven turbulent anisotropy model. The random forest model provided modest improvements to both low and high-blowing ratio film cooling cases along centreline and spanwise distributions. Large cooling effectiveness improvements (up to 82%) were found when compared to the Gradient Diffusion Hypothesis (GDH) model and marginal improvements were shown when compared to the HOGGDH model with its standard coefficient of 0.6.