Tuning of Generalized K-Omega Turbulence Model by Using Adjoint Optimization and Machine Learning for Gas Turbine Combustor Applications

Abstract

Abstract Swirl-stabilized gas turbine combustors often use turbulence modeling through Scale-Resolved Simulations (SRS), like Large Eddy Simulations (LES). However, LES is computationally intense due to its large mesh and small time-step demands. Reynolds Averaged Navier Stokes (RANS) models, especially the Generalized k-omega (GEKO) version, are efficient but may not capture detailed swirl patterns precisely. In previous studies, the Generalized k-omega (GEKO) based RANS model has been used to simulate turbulent flows. The free-coefficients of the GEKO model are tuned to predict the characteristics of potential flow undergoing separation. In this paper, we use the adjoint method and machine learning (ML) to build an argument Neural Network (NN) model for the GEKO coefficient to match the RANS simulation w.r.t. LES simulation. This allows computationally faster simulations of swirling flow compared to LES. An industrial configuration named the DLR PRECCINSTA burner has been used considering a cold non-reacting flow. The objective function defined in this work is based on the linear combination of differences in velocity components, and a Neural Network is trained to obtain an optimized GEKO model via the adjoint method. The results with optimized GEKO agree with LES and experiments. The generalization of the NN model is tested on different flow conditions, including different Reynolds numbers and a reacting flow scenario. The results show significant improvements over the baseline solution as compared to the experiment.