A Deep Learning Framework for Supersonic Turbulent Combustion

Abstract

The rapid simulation of supersonic turbulent combustion is a significant demand in scientific research and engineering applications for hypersonic vehicles. This paper proposes a deep learning framework for fast predicting unsteady turbulent combustion flow fields within the combustor of hypersonic vehicles. Based on convolutional neural networks and recurrent neural networks, this framework extracts spatial distribution characteristics of the flow fields and temporal evolution rules. And we enhance the traditional mean square error loss function by assigning loss weights to different channel data. Numerical simulations are conducted on the model scramjet combustor with various geometric structures to generate the dataset for training, and part of the untrained cases are used to verify the effectiveness. The results show that the proposed model, under different geometric structures, achieves high computational accuracy, with a correlation coefficient between the predicted results and the true values above 0.99. Considering the time cost of data transferring between heterogeneous systems, the model takes only 30 seconds to complete the calculation, representing an acceleration of at least two orders of magnitude compared to computational fluid dynamics. In the future, it can be applied to the rapid prediction of hypersonic vehicle performance and efficiently guide the optimal design of aircraft.