Vector-based neural network turbulent heat flux closures in near-wall cooling jet flows

Near-wall jet cooling flows are difficult to model using industrially appropriate CFD methods such as Reynolds-Averaged Navier–Stokes (RANS) approaches. Previous work has addressed the Reynolds stress closure and simple approaches to improve the turbulent heat flux closure. In this paper, two novel vector-based neural network models have been developed to capture complex turbulent heat flux trends found in near-wall cooling jet flows to improve RANS modelling approaches of these complex flows. The first model features diffusion-based vector inputs which can replicate the early jet development region while the second model which uses a broader vector feature space and captures the turbulent heat flux in the early jet development region and the downstream jet region. The latter model replicated the LES streamwise, normal and spanwise profiles of turbulent heat flux and provided improved results over GDH and HOGGDH turbulent heat flux models. • VBNN models replicated the turbulent heat flux in the coolant jet. • VBNN-8 improves the accuracy of the downstream turbulent heat flux. • Methodology captures direction of the turbulent heat flux vector. • VBNN-8 reduces error of streamwise turbulent heat flux from HOGGDH’s 156% to 6%.