FIELD INVERSION AND MACHINE LEARNING STRATEGIES FOR IMPROVING RANS MODELLING IN TURBOMACHINERY

Turbulence and transition modelling are critical aspects in the prediction of the flow field in turbomachinery. Recently, several research efforts have been devoted to the use of machine learning techniques for improving Reynolds-averaged Navier-Stokes (RANS) models. In this framework, a promising technique is represented by field inversion which requires to find an optimal correction field that minimises the error between numerical predictions and experimental data. In this work, Artificial Neural Networks and Random Forests are investigated as tools to generalise the correction provided by field inversion. An approach to automatically identify the regions where the correction model should be computed is proposed: this improves the fitting and reduces the calls to the model during the predictions. Furthermore, a correction-based weighting of the database is introduced in order to improve the training performances. The potential and the issues of the methods are investigated on a high-lift gas turbine cascade at low Reynolds number.