Spatiotemporal prediction of scramjet combustion based on simpler yet better video prediction neural network

Flame stabilization is critically challenging in the advancement of scramjet engines, as it directly affects their operational efficiency and safety. Consequently, accurately predicting the future combustion state of the scramjet flame is essential for implementing proactive measures to mitigate potential safety risks while enhancing fuel efficiency. In this paper, we report a Simpler yet Better Video Prediction (SimVP) model based on kerosene-fueled scramjet combustion images for spatiotemporal prediction. By establishing a temporal relationship between historical image sequences and future ones, the upcoming combustion state of the flame can be generated. In this work, the predictive capability of the proposed method was tested by varying the number of predicted frames. This study compared typical spatial–temporal neural networks, including Recurrent Neural Network (RNN) based models, Temporal Attention Unit model, and the SimVP network, with the last one demonstrating the best feature capture capability and hence prediction performance. It was able to predict seven future flame frames based on an input sequence of eight frames, while maintaining mean absolute error of 25.6896 ×10−3, mean square error (MSE) of 7.2821 ×10−3, peak signal-to-noise ratio of 23.0342, and structure similarity index measure of 0.9990, reducing MSE by 8.04% and 89.31% compared with ConvLSTM and PredRNN, respectively. The SimVP method reported in this work, due to its superior accuracy and novel architecture, holds promise as a potential approach for spatiotemporal prediction of supersonic turbulent combustion.