Physics-Informed-Neural-Network-Based Fast Prediction Method for Scramjet Nozzle Flow

To enable fast, efficient, and reliable prediction of the nozzle flowfield to assist optimization design and knowledge discovery, this study investigates a dual data- and knowledge-driven model approach for nozzle flowfield prediction. Traditional data-driven flowfield prediction models exhibit limitations in prediction accuracy and physical interpretability, and the skewed distribution characteristics of raw computational fluid dynamics (CFD) data significantly constrain the effectiveness of model training. To address these issues, a structured grid bilinear data enhancement (SGBDE) method is proposed, which effectively improves the distribution characteristics of the CFD data. Simultaneously, a physics-informed neural network (PINN) framework is constructed by incorporating Euler equation constraints and flow physical quantity vector structure constraints, thereby enhancing the model’s generalization capability. By integrating SGBDE and PINN, DE-PINN is constructed. Experimental validation shows that the coefficient of determination [Formula: see text] for physical quantity predictions by DE-PINN exceeds 0.997, significantly outperforming other models. Ablation experiments further validate the individual effectiveness and synergistic interaction of SGBDE and physics-informed constraints, demonstrating excellent application potential, providing new technical support for nozzle optimization, and offering important references for intelligent flowfield prediction.