Physics-Informed Machine Learning for Predicting Ice Particle Impact Dynamics.

Accurately predicting the rebound behavior of ice particles is essential for improving anti-icing and de-icing strategies in transportation, aerospace, and renewable energy systems. This study introduces a data-driven framework using machine learning (ML) to estimate the normal coefficient of restitution (CORₙ) of ice particles across diverse impact conditions. By reformulating five major experimental datasets using Reynolds (Re), Froude (Fr), and Péclet (Pe) numbers, we enable consistent interpretation of heterogeneous data. Among the models tested, the extreme gradient boosting (XGBoost) algorithm achieved the highest accuracy, with a testing R² of 0.93 and a mean absolute error of 5.6%, surpassing traditional physics-based predictions. A physics-informed ML (PIML) variant, combining a log-linear regression baseline with XGBoost-based residual learning, provided improved interpretability and generalization, yielding R² of 0.86. Model analysis highlighted how increasing Re, Fr, and Pe values contribute to greater energy dissipation via viscous, inertial, and thermal effects, respectively. These findings support the development of an accessible design chart delineating elastic to inelastic impact regimes, and the derivation of symbolic expressions further strengthens the model’s utility in engineering practice. This integrated ML–PIML approach offers a predictive and physically grounded tool to guide the design of ice-repellent surfaces and impact-resistant systems.