Rapidly predicting the temperature field of the aero-engine combustion chamber can assist researchers in quickly understanding the combustion state. Although deep learning methods can solve this problem in some situation, the generalization ability of this method still faces challenges. This research proposes a deep learning prediction scheme with the loss function based on high-temperature deviations for the temperature field of an aero-engine combustion chamber. Comparative results indicate that the network models with high-temperature deviation constraints demonstrate significantly better prediction accuracy outside the learning area compared to network models without physical constraints. The results also demonstrate that the involvement of the physical constraint loss function at different stages of network training has a significant impact on the predictive performance of the network model and generalization ability. Compared to network models trained solely on mean square error, the fully-connected network model with mid-to-late stage of physical constraint training improves by 64.5%, and the optimized attention network model improves by 94.0%.
Read full abstract