Abstract
Turbulence modeling plays a crucial role in swirl-stabilized gas turbine combustors, typically relying on scale-resolved simulations (SRS) such as large eddy simulations (LES). However, LES is computationally expensive due to the need for fine mesh resolution and small time-steps required to capture the combustor's large-scale turbulence motion accurately. On the other hand, Reynolds-averaged Navier–Stokes (RANS) models while computationally efficient, lack fidelity in predicting complex flow characteristics such as swirl accurately. In this study, the GEKO model is used for simulating RANS predictions of turbulence in a swirling flow scenario, whereas high-fidelity LES predictions serve as target data enabling field inversion via gradient-based optimization using the Adjoint solver in ansysfluent. Machine learning via neural network (NN) Training is employed to establish correlations between turbulent flow features and optimal GEKO parameters, enabling the trained model's generalization. This approach allows computationally faster simulations of swirling flow using an optimized GEKO model matching the predictions of LES. This methodology is tested on the DLR PRECCINSTA burner considering a nonreacting, isothermal flow scenario. The velocity field variance between LES and GEKO-RANS solutions is defined as the objective function, with the turbulent kinetic energy (TKE) source coefficient serving as the tuning parameter. Results using the optimized GEKO model demonstrate qualitative and quantitative agreement with LES and experiments. The trained neural network (NN) model's generalization is tested on various flow conditions, including six additional Reynolds numbers and a reacting flow scenario, showcasing significant improvements over the baseline model solution. This optimized workflow holds promise for future studies involving different geometries with similar flow fields.
