Machine Learning modelling for nitrocellulose-based propellant ageing

Abstract. Stabilizer depletion is a primary ageing indicator in nitrocellulose-based propellants and constitutes a key surveillance parameter within NATO practice. However, translating accelerated ageing results to realistic storage temperatures remains challenging due to the strong temperature dependence of degradation kinetics and the limited availability of low-temperature data. This study compares two modelling strategies for predicting stabilizer loss in 30 mm- caliber ammunition. Method 1 employs a purely data-driven Gaussian Process Regression (GPR) model trained on accelerated stabilizer measurements, with temporal augmentation and cross-validation for model selection. Method 2 employs a physics-informed hybrid approach in which high-temperature kinetic constants are extrapolated to 16-30°C using an Arrhenius law and first-order decay, thereby generating a thermodynamically consistent degradation surface, which is subsequently emulated by GPR for efficient non-isothermal evaluation. The results show that embedding Arrhenius kinetics within a machine learning framework enhances predictive robustness, physical interpretability, and scientific defensibility for operational shelf-life assessment.