Physics-informed machine learning model for prediction of long-rod penetration depth in a semi-infinite target

Traditional data-driven approaches to long-rod penetration modeling incorporate significant expert bias through handcrafted analytical formulas. In this work, we investigate machine learning as an alternative to build a more flexible, less biased model. Published data from approximately 900 experiments were used to train an artificial neural network to predict the depth of penetration into a semi-infinite target. On the basis of empirical parameter studies, we optimized the neural network architecture and regularization technique for this task. To aid the learning process, we incorporated into the machine learning model general physical principles, such as formalized in the Buckingham Pi theorem. The resulting physics-informed machine learning model shows good performance in generalization to new combinations of penetrator and target materials. As exemplary model application, we accurately postdict experimental results on the influence of target hardness fluctuations on penetration depth. [Display omitted] •Machine learning is applied in the context of long-rod penetration mechanics.•Artificial neural network predicts depth of penetration in semi-infinite target.•Buckingham Pi theorem is incorporated into a physics-informed machine learning model.•We postdict variations in penetration depth due to target hardness fluctuations.•Data-based model offers less biased alternative to manual data parametrization.