Abstract
This paper presents a complete and general machine-learning assisted optimization framework for the Generalized k-ω (GEKO) turbulence model based on experimental measurements. The optimization framework is applied to the use case of a turbocharger radial compressor for three different operating conditions including design and surge conditions. The first optimization of the GEKO turbulence model is performed w.r.t measured global thermodynamic quantities, including the isentropic efficiency and pressure ratio. A second optimization is performed w.r.t local experimental flow quantities, i.e. wall pressure distribution along the diffuser vane. The testing errors of the machine-learning models are compared and analyzed, and the best performing algorithms are Support Vector Regression (SVR) and Artificial Neural Network (ANN) for modeling the response surface of the GEKO parameters, based on which the turbulence model is optimized by Genetic Algorithm (GA). The aforementioned 2 optimizations result in an equivalent GEKO parameter set, where the optimized GEKO model agrees well with the measurements and the more computationally expensive Explicit-Algebraic Reynold Stress Model (EARSM), and outperforms the Shear Stress Transport (SST) in terms of relative error, and the prediction of onset of surge and flow separation. The optimized GEKO model also generalizes well to different operating conditions as well as another radial compressor with a different diffuser geometry. In addition, a data-based approach is used to quantify the influence of the six GEKO parameters on different flow phenomena in the radial compressor. The result of the data-based approach aligns well with heuristic loss mechanism in turbomachinery.
