Aero-thermal coupling optimization for a turbine blade with lamilloy cooling structure based on machine learning

This paper applies machine learning to optimize a lamilloy cooling structure in a turbine, focusing on minimizing the high-temperature region on the turbine blade surface. Seven geometric parameters were used to define the cooling structure and 200 parameter combinations were generated via Latin hypercube sampling. An in-house parameterization procedure produced 189 valid geometric models, and the high-temperature regions were evaluated using numerical simulations. A BP neural network with a learning rate of 0.00001 and five hidden layers was developed to map the parameters to the high-temperature area, achieving high prediction accuracy. The model reduced losses to 0.0005 on the training set and 0.0085 on the test set. Optimization based on the trained neural network showed that the percentage of high-temperature area on the blade surface decreased from 0.13461 to 0.0473 and 0.0642 when partial and full geometric parameters were used as optimization variables, respectively, verifying the model accuracy.