Fast performance prediction and field reconstruction of gas turbine using supervised graph learning approaches

Accurately and rapidly predicting the multi-conditions characteristics of turbines is fundamental for realizing efficient energy conversion and optimal layout schemes. Based on supervised graph learning approaches, this work is dedicated to establishing a fast multidisciplinary prediction model for gas turbines. The research target is a gas turbine blade with complex cooling channels. An Aerodynamic Strength Prediction Graph neural network (ASP-GNN) is proposed to predict the aerodynamic-strength characteristics and temperature field under different boundary conditions. The superiority of our approach is demonstrated by prediction precision and time cost. The generalizability of the network is also investigated by adopting different training set sizes, and the ASP-GNN can achieve satisfactory accuracy with a limited number of training samples. Based on the established model, the effects of various boundary conditions on aerodynamic and strength performance are quantified. The unsteady characteristic of performance and temperature field are also obtained conveniently. The proposed model could serve as a fast analysis approach to aid the design and analysis of turbomachinery. It may relieve the workload of numerical simulations for complex engineering analysis.