A predictive surrogate model for heat transfer of an impinging jet on a concave surface

This paper aims to comprehensively investigate the efficacy of model order reduction and deep learning techniques in predicting heat transfer of pulsatile impinging jets on a concave surface. We introduce two predictive approaches: one employing a Fast Fourier Transform-augmented Artificial Neural Network for predicting the area-averaged Nusselt number under constant-frequency jet scenarios, and another comparing the performance of LSTM and Transformer neural networks for random-frequency jets. Results indicate that the Transformer significantly outperforms the LSTM, achieving higher accuracy and robustness in predicting Nusselt numbers — covering up to 50% of the cycle with precision, whereas the LSTM covers only 20% with greater error margins. Additionally, the integration of Proper Orthogonal Decomposition with Transformer networks yields a novel strategy for predicting local Nusselt number distributions. This method significantly reduces computational complexity while maintaining high accuracy, with a maximum prediction error of 5%. These findings demonstrate the efficacy of advanced deep learning techniques for temporal prediction of the Nusselt number on complex surfaces, suggesting further applicability in broader fluid dynamics and heat transfer analyses.