Hierarchical Attention Modeling Method for Complex Combustion System Simulation in Scramjets

Strong coupling and small reaction zones lead to extremely high computational cost for high-fidelity simulation in supersonic combustion design. Accurate full-field predictions of temperature, pressure, and species provide an efficient surrogate for monitoring and performance assessment. This paper proposes a Hierarchical Residual Attention Network (HRANet) that inputs flow coordinates, time, Mach number, and fuel type to output full-field distributions of temperature, pressure, density, and species mass fractions. HRANet uses a shared residual attention backbone with attention mechanisms to focus on sharp gradients like flame fronts. Physical fields are grouped into stable, coupled, and extreme categories, each with a dedicated subnetwork; the extreme group incorporates multiscale convolutions to capture both local and global features. A hierarchical loss function, adapted to statistical distributions, enhances reconstruction of weak-signal fields. Validated on a combustion data set covering kerosene, ethylene, hydrogen, and multiple equivalence ratios, HRANet outperforms all benchmarks, achieving peak signal-to-noise ratio less than or equal to 29.8 dB and coefficient of determination [Formula: see text], with density [Formula: see text] improved by over 32.4%. The method enables cost-effective high-accuracy reconstruction for extreme combustion multiphysics, with engineering benefits in accelerating simulations and supporting intelligent monitoring.