Optimized neural network for supersonic isolator flow field prediction incorporating prior information and attention mechanisms

The rapid and accurate prediction of the flow field during supersonic isolator operation is crucial. Deep learning-based pressure monitoring during operation is an effective method for flow field prediction. A supersonic isolator flow field dataset was produced for a ground-based experiment with a variable incoming Mach number and back pressure. An approach for predicting the future flow field based on isolator pressure monitoring was proposed. A flow field prediction model incorporating long short-term memory, temporal convolutional network, and convolutional block attention module structures has been proposed. The performance of the proposed model was analyzed and compared with those of other time-series neural networks for flow field prediction. The location of the shock train leading edge was introduced as a priori information to enhance the model prediction performance. The impact of the weights associated with the a priori information in network training on the performance of the flow field prediction model was analyzed and discussed. This study presents an optimization scheme for neural network flow field prediction models specifically tailored for the supersonic isolator flow field prediction problem.