Improvements in the Prediction of Steady and Unsteady Transition and Mixing in Low Pressure Turbines by Means of Machine-Learnt Closures

In this paper novel machine-learnt transition and turbulence models are applied to the prediction of boundary-layer transition and wake-mixing in a Low-Pressure Turbine (LPT) cascade, including

unsteady inflow cases with incoming wakes. A Laminar Kinetic Energy (LKE) transport approach is employed as baseline transition framework, and the machine learning algorithm was exploited to reformulate source terms. In addition to that, explicit algebraic Reynolds stress models were trained to improve wake mixing predictions. The application of the machine learnt EARSMs takes advantage of a zonal strategy based on a newly developed sensing function that allows automated wake demarcation. This work compares the performance of several approaches that are based on the application of improved transition models without machine learnt EARSMs, baseline transition model with EARSMs trained for improved wake mixing, and new transition and turbulence closures that are simultaneously developed in a fully coupled way and aimed at improving both transition and wake mixing predictions in LPTs. The investigation is carried out on a cascade of industrial footprint, representative of modern LPT bladings. First of all, the various modelling frameworks are evaluated, without bar wakes (steady conditions), over a wide range of Reynolds number values. Most of the considered conditions are far from the training environment of the proposed closures, and such an extended simulation campaign is intended to assess their generality and robustness. URANS analyses have also been carried out to investigate the predictive capabilities of machine-learnt closures in the presence of incoming bar wakes.

RANS/URANS results are scrutinized against LES calculations and experimental data. It will be demonstrated that machine learnt EARSMs are capable of producing realistic wake mixing when simply applied on top of the baseline transition models. However, the most accurate predictions will be shown to be provided by fully integrated machine-learnt transition/turbulence closures.