Application of Machine Learning Based Meta Models for Predicting Film Cooling Effectiveness in Gas Turbine Blades

Abstract

Abstract In Large Gas Turbines, turbine components in particular blades and vanes operate at significantly high temperatures. As a result, cooling of these components becomes one of the prime objectives during the design phase. There are different cutting-edge internal cooling techniques such as convection cooling - internal ribbed surface heat transfer only, serpentine passage, impingement cooling, film cooling and combination of all these methods are used to meet the heat transfer design requirements. Among all methods, film cooling plays a vital role in the improvement of heat transfer and is extensively applied to ensure thermal protection of hot gas path components especially blades and vanes. A precise estimation of film cooling effectiveness is critical for evaluating external heat transfer and metal temperature. Conventional approach based on heat transfer correlations and finite element method with varied film cooling scheme is highly resource intensive which involves setting up different cooling schemes. The model setup process includes geometrical modeling, meshing of model, boundary conditions application and performing thermal simulation to achieve optimal design. This constrained the number of film cooling hole configuration which could be evaluated, resulting in sub-optimal design in terms of cooling effectiveness. In this work Machine learning approach is explored for establishing optimal film cooling hole configuration with minimal time and efforts and can be applied during early design phase. Film cooling effectiveness is influenced by several factors like coolant & mainstream conditions, cooling hole arrangement and other geometrical parameters such as cooling hole diameter, inclination, pitch, breakout, coverage, and location on the airfoil geometry etc. In current scope cooling hole diameter, pitch, cooling hole count, and surface angle is used as input feature to the Machine Learning model and cooling effectiveness as target variable. The input to the Machine learning model is derived from physics-based heat transfer calculations. A comparative study is introduced between linear regression, K-nearest neighbour, random forest, and gradient boost Machine learning methods using root mean square values as an evaluation metrics. Optimal hyperparameter were calculated for each model using grid search method. It was observed that gradient boost method shows minimum error, additionally cooling hole count is the most significant factor in predicting cooling effectiveness.