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. A tensor-basis neural network 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 are employed directly into RANS solver, augmenting a k-omega 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. Underpredictions of the turbulent kinetic energy field, modeled with the SST transport equation, were 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 advance the current state-of-the-art in RANS modeling of film and effusion cooling flows.