Prediction of 2D film cooling effectiveness distribution: A generative neural network with physical prior knowledge

Abstract

Film cooling is an essential thermal protection technology that directly influences the performance of hot-end components. Its effectiveness affects combustion efficiency and significantly influences pollutant and carbon emissions during combustion. Consequently, the rapid design and evaluation of cooling schemes have become critical research priorities. Traditional neural network prediction models, however, demand large datasets, with data acquisition costs often being high. This study integrates physically meaningful prior knowledge with image encoding and decoding modules that utilize multi-head attention mechanisms. The goal is to enhance the prediction accuracy of the two-dimensional distribution of film cooling effectiveness ($\eta$) with limited sample sizes. Furthermore, a highly reliable PSP measurement system was developed to substitute for sample sets generated by CFD simulations. The results indicate that, compared to the traditional model with prediction errors for $\eta$ and non-uniformity ($\sigma$) exceeding 50 %, the proposed model can control the prediction accuracy within the range of 5 % to 15 %. Furthermore, the integration of encoding and decoding modules with a multi-head attention mechanism allows the model to excel in predicting local distributions while also improving its generalization ability. The gradient-based sensitivity analysis on the input structural parameters revealed that three factors—spacing P, exit width, and inlet-to-outlet area ratio—exhibit more pronounced effects on $\eta$.