Data driven turbulence modeling in turbomachinery — An applicability study

The machine learning (ML) approaches have been introduced in Reynolds-Averaged Navier–Stokes (RANS) modeling in recent years. These ML-RANS models are usually trained from a training flow database, and are employed in the simulations of other flow cases. The training flows and real simulations can have different underlying physics. When the training flows have a typical physical constraint but real simulations do not have, the applicability of ML-RANS models should be evaluated. The present paper then aims at explaining the underlying physics and expansibility of a recent ML-RANS approach (Liu et al., 2021). A database of transitional channel flow is first used to estimate the flow states and data constraints, and to clarify the calculation methods for the input features. An industrial low pressure turbine (LPT) configuration is then calculated to test the applicability of the applied ML-RANS model. Analysis on the expansibility of constraints shows that the non-equilibrium turbulence and transitional flow cannot be correctly calculated by using a ML-RANS model learned from databases of equilibrium turbulence. •A-priori performance of the machine learning model is investigated on a LES database•A-posteriori analyses are performed on an in-house cascade test configuration•Prediction capability of the machine learning model is limited by the training data