Data-driven turbulence anisotropy in film and effusion cooling flows

Film and effusion cooling flows contain complex flow that classical Reynolds-Averaged Navier Stokes (RANS) models struggle to capture. ATensor-BasisNeuralNetwork (TBNN) is employed to provide an anisotropic model that can reproduce the Reynolds stress fields of Large-Eddy Simulations (LES). High-quality LES datasets are used to train, validate and test a neural network model. A priori results show the model can reproduce the Reynolds stress field on a cooling case not present in the model’s training. The neural networks is employed directly into RANS solver, augmenting a 𝑘-𝜔 Shear Stress Transport (SST) model, with conditioning applied. The model provided improvements to Reynolds stress, velocity and temperature fields in cases not used to train the model, including a multi-hole case that differs from the single-hole geometry used to train the case. Under predictions of the turbulent kinetic energy field, modelled with the SST transport equation, was found to lead to underpredictions in the neural network produced Reynolds stresses. Correcting this with the LES resolved turbulent kinetic energy provided further agreement. The method found significant improvements to the surface cooling results that advances the current state-of-the-art in RANS modelling of film and effusion cooling flows.