Physics-informed machine learning based RANS turbulence modeling convection heat transfer of supercritical pressure fluid

Abstract

Solving flow problems based on the Reynolds-averaged Navier-Stokes equations is a dominant method in terms of efficiency and accuracy for design and analysis of engineering applications, while the performance of RANS models in predicting the turbulent heat transfer of supercritical pressure fluid could be severely poor. To a great extent, the models are plagued by imperfect closures on higher order turbulence quantities with strongly varying thermal physical properties. This paper reports on an alternative approach to model the turbulent heat transfer at supercritical pressure, including the direct turbulent production models and the indirect transport closures using deep neural networks (DNN). This paper presents a method of modifying turbulence models of supercritical pressure fluid from high-fidelity simulation (DNS) data. An iterative DNS-DNN-RANS framework is proposed to establish explicit closures for turbulent momentum diffusion and turbulent thermal diffusion of turbulent heat transfer at supercritical pressure. Priori physics knowledge, feature engineering strategy and establish existing constitutive theory are involved to embed the special characteristics of supercritical pressure fluid into the turbulent closures to establish a proper regression system for machine learning (ML) algorithm. By embedding the explicit ML models, the low Reynolds number k-ε model was modified and was trained under a cross-case strategy with abundant DNS data. The modified model was successfully validated against DNS and experimental data for upward pipe flows, in which wall temperature were satisfactorily reproduced. The posterior simulations showed that the modified ML-KTVT model not only has a good performance in predicting the convection heat transfer of supercritical pressure fluid but also performs a favorable reproducibility of turbulence development in heat transfer deterioration cases.