Physics-informed machine learning for predicting the ballistic limit of whipple shields

• Artificial neural network (ANN) based on ballistic limit equation (BLE) structure demonstrated. • Gaussian process (GP) with BLE as prior mean demonstrated. • Incorporating physics-based BLE has little effect on ML model classification accuracy. • Incorporating BLE allows ML models to extrapolate to conditions outside the training dataset. • Physics-informed ML models can overcome some limitations of existing data-drive approaches. Data driven machine learning (ML) models can provide improved accuracy over semi-analytical ballistic limit equations (BLEs) for predicting the outcome of space debris impacts on spacecraft structures. However, they should not be applied beyond the scope of their training data which limits their utilisation in mission risk assessments. We develop and demonstrate two approaches for incorporating physics knowledge, in the form of existing BLEs, into ML models to mitigate this limitation. The resulting physics-informed models provide modestly improved classification accuracy when applied on a database of experimental records as well as improved agreement with BLEs when applied outside the scope of the training dataset, compared to previous data-driven ML models.