Physics-Informed Machine Learning for Predicting the Ballistic Limit of Whipple Shields

Abstract Introduction Ballistic limit equations (BLEs) define the performance of a spacecraft structure or component impacted by a space debris particle, typically in terms of a critical projectile diameter for a given impact velocity and obliquity. BLEs are typically based on analytical considerations, e.g., kinetic energy scaling, etc., but utilise varying degrees of empiricism to improve agreement with experimental data. When applied to predict the binary pass/fail outcome of hypervelocity impact experiments, including those on which they were empirically fit, they are accurate approximately 70-75% of the time (according to [1]). In [2] several machine learning (ML) classification models were developed for predicting the ballistic limit of Whipple shields, achieving improved test accuracies of 77-85%. However, data-driven ML models are not generalisable, i.e., they are only valid only for application on conditions encompassed in their training data. This is a major limitation. Ideally, we desire a model that can achieve the improved accuracy of data-driven ML models, when we have experimental data for training, and that is generalisable to any hypervelocity impact conditions that a shield may be expected to encounter in orbit when we don’t.