A random forest machine learning in turbulence closure modeling for complex flows and heat transfer based on the non-equilibrium turbulence assumption

Turbulence models generally require specific corrections or optimization of turbulence constants for predicting the complex flows and heat transfer accurately. However, the high-fidelity methods demand extensive computational costs for solving these complex phenomena. To address this issue, a random forest machine learning driven turbulence model is proposed, based on the non-equilibrium turbulence assumption and in accordance with the traditional turbulence models. The proposed framework adjusts the energy production and dissipation to achieve the non-equilibrium turbulence properties without learning the Reynolds stresses, unlike other machine learning methods. This key feature allows the model to utilize low- and high-fidelity data, broadening its applicability and solving stability. The proposed method is trained on fourteen cases, including the laminar-turbulence transition flows, the jet impingement, the swirling flow, and the reattachment flow. Many unseen cases with different physics are used to evaluate the performance of the above method in terms of prediction accuracy and solving properties. Also, a backward-facing step is estimated for the treatment of high-fidelity data. The results show that the proposed method has the potential to get more accurate results than the reference model not only in the velocity field but also in the heat transfer rate. Additionally, the proposed method consistently produces convergent and robust results, even with changes of geometries and operating conditions.