A novel data-driven method for augmenting turbulence modeling for unsteady cavitating flows

TL;DR

This paper introduces a data-driven method integrating Gene-Expression Programming with RANS to improve turbulence modeling in unsteady cavitating flows, addressing limitations of traditional models.

Contribution

It presents a novel GEP-based correction technique for RANS models, enhancing their accuracy in simulating cavitating flows without requiring continuous feedback.

Findings

GEP-CFD outperforms baseline RANS in capturing flow features.

The method effectively models Reynolds shear stress and turbulent kinetic energy.

Compared to linear regression, GEP provides more complex and accurate corrections.

Abstract

Cavitation is a highly turbulent, multi-phase flow phenomenon that manifests in the form of vapor cavities as a result of a sudden drop in the liquid pressure. The phenomenon has been observed as widely detrimental in hydraulic and marine applications like propellers and pumps, while also accelerating the sub-processes involved for bio-diesel production. Modelling cavitating flows has been extremely challenging owing to the resulting flow unsteadiness, phase transition and the cavitation-turbulence interaction. Standard methods to model turbulence like Reynolds-Averaged Navier-Stokes equations (RANS) and hybrid RANS-Large Eddy Simulations (LES) models are unable to reproduce the local turbulence dynamics observed in cavitating flow experiments. In an effort to facilitate the development of accurate RANS modelling procedures for cavitating flows, a data-driven approach devised by integrating Gene-Expression Programming (GEP) into a traditional RANS approach, without continuous RANS model feedback is proposed. The GEP algorithm is employed to obtain an additional corrective term, composed of flow variables derived directly from a baseline URANS cavitating flow calculation to be fit directly in the Boussinesq approximation. In addition, an optimizer based on the BFGS algorithm is employed on top of the GEP algorithm to ensure the solution is independent of the uncertainties associated with GEP. To evaluate the complexity of the model, we compare the approach with a more elementary linear regression technique. The GEP-CFD method is then used to train two expressions, for the Reynolds shear stress and the Turbulent Kinetic Energy simultaneously. The GEP-CFD approach is able to produce the flow features not captured by baseline RANS simulations in several areas, outperforming the RANS simulations and thus projects as a promising approach to augment numerical simulations for cavitating flows.