An iterative machine-learning framework for RANS turbulence modeling

Iterative ML-RANS framework incorporated with conventional turbulence transport equations.•A built-in reproducibility of the training cases is ensured.•NN is trained with extended independent variables to overcome multi-valued issue.•Favorable interpolation capability is achieved by cross-case training strategy.•Improved predictions on separated flow with training data only from channel flow. Machine-learning (ML) techniques provide a new and encouraging perspective for constructing turbulence models for Reynolds-averaged Navier-Stokes (RANS) simulations. In this study, an iterative ML-RANS computational framework is proposed that combines an ML algorithm with transport equations of a conventional turbulence model. This framework maintains a consistent procedure for obtaining the input features of an ML model in both the training and predicting stages, ensuring a built-in reproducibility. The effective form of the closure term is discussed to determine suitable target variables for the ML algorithm, and the multi-valued problem of existing constitutive theory is studied to establish a proper regression system for ML algorithms. The developed ML model is trained under a cross-case strategy with data from turbulent channel flows at three Reynolds numbers, and a posteriori simulations of channel flows show that the framework is able to predict both the mean flow field and turbulent variables accurately. Interpolation tests for the channel flow show the proposed framework can reliably predict flow features that lie between the minimum and maximum Reynolds numbers associated with the training data. A further test of the ML model in a flow over periodic hills also demonstrates improved performance compared with the k-ω SST model, even though the model is only trained with planar channel flow data.