NeuroFire: Application of neural networks for multi-scale combustion performance simulation in solid composite propellants

This study presents NeuroFire, a neural network-based chemical solver framework for multi-dimensional combustion modeling of solid composite propellants. The framework integrates a physics-informed neural operator with adaptive boundary sampling to solve stiff chemical kinetics, overcoming computational bottlenecks in traditional ODE solvers. By coupling a multilayer perceptron trained on 200,000 chemical states with a novel loss-weighted sampling strategy, NeuroFire achieves 92 % accuracy in species prediction while reducing computation time by 90 % compared to conventional CVODE methods. The Taichi-powered implementation enables cross-platform deployment across 0D–2D combustion scenarios, validated through: (1) Zero-dimensional reactor simulations showing <5 % deviation from Cantera solutions, (2) 1D steady-state predictions matching experimental burning rates within 8 % error across 1–10 MPa, and (3) 2D simulations revealing HMX particle size-burning rate positive correlation contrasting AP's inverse relationship, providing critical insights for propellant formulation. NeuroFire's capability to predict ignition delays within 15 % of experimental values under 55–100 W laser ignition further demonstrates its versatility in combustion engineering applications.