Title:An Iterative Machine-Learning Framework for Turbulence Modeling in RANS

Abstract:Machine-learning (ML) techniques provide a new perspective for constructing turbulence models for Reynolds-averaged Navier-Stokes (RANS) simulations. Designed with advanced fitting ability, they can increase the accuracy of the turbulence models given enough information from high-fidelity datasets is provided. In this study, an iterative ML-RANS computational framework is proposed, that combines the ML algorithm and transport equations of a conventional turbulence model built on empirical knowledge. This framework could maintain a consistent procedure to obtain the input features for ML models in both the training and predicting stages. The effective form of the closure term is discussed to explain how to determine the target variables for the ML algorithm, which ensures a mean flow solution of RANS equations free of amplified error. The inherent multi-valued problem of the existing constitutive theory is studied to establish a proper regression system for ML algorithms. From the same input features, an accurate closure can be obtained through ML model, and a flow field similar to the high-fidelity datasets can be obtained based on such closure so that a built-in reproducibility for the training cases can be achieved. It is demonstrated that the framework can deal with a cross-case training strategy with data from turbulent channel flows at different Reynolds numbers. A posteriori simulations of channel flows show that the framework is able to predict both the mean flow field and turbulent variables accurately. Applied to the flow over periodic hills a better result than for a conventional first order turbulence model is obtained, indicating a promising prediction capability of the developed ML-RANS model for a recirculating flow even though the model is trained with planar channel flow data