Framework for Machine Learning-Based Turbulence Modeling to Accurately Predict the Complex Flows in Gas Turbines

Abstract Flow in turbomachinery such as industrial gas turbines is highly complex and exhibits different characteristics for different operating conditions within a duty cycle or performance range. This poses a challenge to predict the performance of the machine accurately and efficiently. Computational Fluid Dynamics (CFD) has been successfully used for aerothermal analysis of turbomachinery successfully. An important aspect of CFD analysis is setting best practice for turbulence models for a given range of the machine especially in a RANS (Reynolds Averaged Navier-Stokes) framework. This is usually a manual process and can be very time consuming based on the complexity of the flow. The ability in many of these turbulence models to adapt the model coefficients to a specific application, such as a gas turbine, is limited. A novel approach is presented in this paper that uses a unique turbulence modeling paradigm and machine learning to accurately and efficiently predict complex flows over a wide range. The new Generalized k-ω (GEKO) turbulence model is unique in that it exposes key and selective coefficients that control the behavior of different flows that may be encountered in gas turbines, such as free shear flow and near-wall flow. This paper provides a generic framework to optimize these key coefficients via neural network-based machine learning algorithm while leveraging available experimental or benchmark data. This approach is applied to an industrial gas turbine exhaust diffuser. Comparisons of the flow predictions using this approach are done against the experimental data for design/off-design conditions.