Cold spraying comprises complex multiscale interactions from particle flight, through coating formation, to final functional performance. However, limited experimental data and the difficulty in measuring transient particle states hinder the development of reliable prediction models. Accordingly, we propose a hierarchical prediction framework for hydrogen evolution reaction (HER) cathode coatings, combining a physics-informed neural network (PINN) and a convolutional neural network (CNN). The PINN embeds momentum- and heat-transfer equations to predict particle impact velocity and temperature, which serve as intermediate physical descriptors. The CNN further establishes nonlinear mappings between these descriptors and coating structural characteristics, elemental composition, and HER electrocatalytic performance. The PINN achieves average R² values of 0.8717 and 0.8026 for velocity and temperature prediction, demonstrating improved stability compared with conventional machine learning models. The CNN achieves reliable predictions for porosity (R²=0.8339), 3D surface area (0.7665), Ni mass ratio (0.7545), overpotential (0.7660), and Tafel slope (0.7162). Subsequently, ablation experiments and SHAP analysis are conducted to identify momentum transfer, heat exchange, and particle-state evolution as the dominant factors governing coating formation and HER electrocatalytic performance. This work provides a physics-informed hierarchical modeling strategy for interpretable prediction and optimization of complex small-sample manufacturing processes.
 

Cold spray,PINN,Hierarchical prediction framework,HER electrocatalytic coating,Interpretable machine learning