Are random forests better suited than neural networks to augment RANS turbulence models?

Abstract

Machine-learning (ML) techniques have bloomed in recent years, especially in fluid mechanics applications. In this paper, we trained, validated and compared two types of ML-based models to augment Reynolds-averaged Navier–Stokes (RANS) simulations. The methodology was tested in a series of flows around bumps, characterized by different levels of flow separation and curvatures. Initially, the ML-based models were trained in three configurations presenting attached flow, small and moderate separation and tested in two configurations presenting incipient and large separation. The output quantity of the machine-learning model is the turbulent viscosity as done in Volpiani et al. (2022). The new models based on artificial neural networks (NN) and random forest (RF) improved the results if compared to the baseline Spalart–Allmaras model, in terms of velocity field and skin-friction profiles. We noted that NN has better extrapolation properties than RF, but the skin-friction distribution can present small oscillations when using certain input features. These oscillations can be reduced if the RF model is employed. One major advantages of RF is that raw quantities can be given as input features, avoiding normalization issues (such as division by zero) and allowing a larger number of universal inputs. At the end, we propose a mixed NN-RF model that combines the strengths of each method and, as a result, improves considerably the RANS prediction capability, even for a case with strong separation where the Boussinesq hypothesis (and therefore the eddy-viscosity assumption) lacks accuracy.