Multi-pressure modalities integrated artificial intelligence model for supersonic isolator flow field prediction with attention mechanisms

The complex flow dynamics in supersonic isolators, under varying Mach numbers and backpressure conditions, are presented with significant challenges for accurate flow field prediction. These challenges are driven by intricate shock - boundary layer interactions, transient shock train motions, and multiscale flow structures. To address these challenges, a neural network framework on artificial intelligence (AI) is proposed in this study. Spatiotemporal pressure characteristics are synergistically integrated through multimodal decomposition and attention - based fusion within this framework. Temporal and spatial modalities, derived from ground - test pressure data, are processed via parallel network branches that combine Long Short - Term Memory and Temporal Convolutional Networks. Attention mechanisms are utilized to dynamically prioritize critical flow features across scales. Compared to conventional single - modality methods, the proposed approach is demonstrated to have enhanced capability in capturing shock system evolution and boundary layer dynamics. By effectively reconciling localized pressure fluctuations with global flow patterns, the framework is shown to exhibit stable predictive performance across diverse operating regimes, providing a promising data - driven paradigm for compressible flow modeling. This work advances the development of intelligent prediction tools for supersonic propulsion systems and provides valuable insights for flow control and isolator optimization under validation conditions.