Enhancing the accuracy of transition models for gas turbine applications through data-driven approaches

Separated-flow transition is a very popular phenomenon in gas turbines, especially in low-pressure turbines (LPTs). Low-fidelity simulations are often used in gas turbine design. However, they cannot predict separated-flow transition accurately. To improve the prediction of separated-flow transition for LPTs, empirical relations that are derived for transition prediction need to be significantly modified. To achieve this, machine learning approaches are used to investigate a large number of functional forms using computational-fluid-dynamics-driven gene expression programming. These functional forms are investigated using a multi-expression multi-objective algorithm in terms of separation onset, transition onset, separation bubble length, wall shear stress, and pressure coefficient. The models generated after 177 generations show significant improvements over the baseline result in terms of the aforementioned parameters. All of the models developed improve wall shear stress prediction by 40%-70% over the baseline laminar kinetic energy model. This method has immense potential for improving boundary layer transition prediction for gas turbine applications across several geometries and operating conditions.