Machine-learning strategies for transition/turbulence modelling for Low-Pressure Turbines with unsteady inflow conditions
| Title | Machine-learning strategies for transition/turbulence modelling for Low-Pressure Turbines with unsteady inflow conditions |
| Publication Type | Journal Article |
| Year of Publication | Submitted |
| Authors | Gu Y, Fang Y, Akolekar HD, Pacciani R, Marconcini M, Ooi ASH, Sandberg RD |
| Journal | ASME J Turbomach |
| Keywords | Data-Driven Computational Methods, Low-Pressure Turbines, Multi- objective Optimisation, nsteady Aerodynamics |
| Abstract | Unsteady flow behaviour induced by wake-blade interactions is crucial for the operational efficiency, aerodynamic stability, and fatigue life of low-pressure turbines (LPTs), and yet remains challenging to capture with (unsteady) Reynolds-averaged Navier–Stokes (U)(RANS) calculations. This paper focuses on a more reliable estimation of unsteady wake-induced losses, arising primarily from wake mixing and boundary-layer transition under periodic disturbances. To achieve this, the CFD-driven training framework is, for the first time, tailored to LPT flow unsteadiness, enabling revisions of both the transition and turbulence models. Firstly, new physics-related features are incorporated into turbulence closure formulations for automated wake-region differentiation, and a new transition-model output is introduced to capture unsteady wake-induced transition. Secondly, model evaluation metrics are supplemented with phase-lock averaged cost functions to ensure consistent improvement throughout the entire unsteady cycle. The integration of transition and turbulence modelling components is achieved in a sequential manner. A comprehensive assessment of both transition and turbulence models is performed using metrics of time-averaged and phase-lock averaged flow features along with secondary statistics, all of which demonstrate solid improvements over the baseline. Detailed model interpretation is also presented to reveal underlying physical insights. Moreover, a-posteriori validation with the machine-learnt transition–turbulence model on different incoming wake frequencies exhibits robust performance, significantly improving the prediction of both wake losses and transition behaviour, not only in a mean sense but also for individual phases. This study highlights the potential of RANS-model development for unsteady multi-stage turbomachinery configurations and provides physical insights into wake-induced unsteadiness in LPTs. |
| Refereed Designation | Refereed |