Interpretable physics-informed Kolmogorov–Arnold networks for linking casting-induced segregation to burning-rate non-uniformity in composite solid propellants

Burning-rate stability in composite solid propellants is influenced by casting-induced particle segregation and the resulting spatial variability in local combustion conditions. To quantitatively resolve the "process variables (casting speed, vacuum level, and slurry temperature) – structural metrics (solid particle distribution) – combustion performance (uniformity)" chain, this study presents a collaborative framework integrating numerical simulations, experiments, and deep learning. First, multiphase flow casting—accounting for the multiphase dynamics of ammonium perchlorate (AP) and cyclotetramethylene tetranitramine (HMX) particles—was simulated via CFD to capture the post-casting particle distribution, followed by burning rate experiments at spatially distinct locations of the cured propellant grain. To bridge these, a Physics-Informed Kolmogorov-Arnold Network (PI-KAN) was constructed to establish the quantitative correlations among process parameters, particle distribution, and the burning rate, utilizing the SHapley Additive exPlanations (SHAP) framework for interpretability. CFD analysis identifies casting speed as the dominant factor regulating uniformity; specifically, increasing the speed within 30–45 kg/h significantly mitigates segregation, narrowing the burning rate variation from 0.566 mm/s to 0.305 mm/s. Crucially, SHAP analysis suggests a model-predicted casting-speed threshold of approximately 42 kg/h, providing an interpretable quantitative reference for process adjustment within the investigated system.