Abstract: The development of scramjet engines relies heavily on accurate simulation of the complex flow fields within the combustor; however, traditional numerical simulation methods are computationally expensive and time-consuming, severely limiting the efficiency of design iteration. Data-driven approaches for rapid flow-field prediction have therefore become an important research direction, yet existing deep learning–based prediction models often rely on spatiotemporal modeling modules such as ConvGRU and ConvLSTM, which are computationally inefficient and difficult to parallelize during training, and which also struggle to effectively distinguish and process structurally complex regions and smooth regions in the flow field. To address these challenges, this paper proposes a novel deep learning model based on a spatiotemporal attention mechanism, which incorporates a content-aware hybrid module to dynamically allocate computational pathways according to the local complexity of flow-field features: self-attention mechanisms are employed to capture global dependencies in strongly nonlinear regions, while lightweight convolutional operations are used for smooth regions, thereby enabling efficient and high-accuracy prediction of no-constant flow field in scramjet combustors. Experimental results demonstrate that, compared to the ConvGRU-based prediction model, the proposed model achieves better accuracy and stability in both single-step and multi-step prediction tasks. The overall mean squared error in single-step and recursive predictions is reduced by 45% and 49%, respectively`, significantly reduced temporal error accumulation, and the capability to forecast flow-field evolution over physical time horizons from 6 ms to 20 ms within 1 s of computation, yielding computational efficiency improvements of several orders of magnitude over traditional CFD methods.