Application of machine-learning approaches with physics-informed oversampling to determine ballistic limit curves

The proposed study investigates the application of physics-informed oversampling in machine-learning models to approximate ballistic curves and ballistic limit velocities that represent the resistance of targets made from S355 steel under impacts of 7.62 × 51 mm P80 (0.308 Win) small-caliber projectiles. Through a detailed terminal ballistic experimental investigation, ballistic limit velocities, along with the corresponding limit curves, were determined for the four target thicknesses, i.e. the plates of the thickness of 20 mm, 16 mm, 8 mm, and 5 mm were used. The experimental results concerning the four impact configurations were further used to validate the finite element analysis, resulting in the identification of an additional set of eleven ballistic limit curves and ballistic limit velocities for a total of fifteen configurations of the S255 steel plates, with thicknesses ranging from 4~mm to 18~mm. Finite element simulations were used to complement the experimental data and further characterize the effects of varying plate thickness on their protection level. Several machine-learning approaches, including deep neural networks (DNN), Random Forests (RF), Support Vector Machines (SVM), and XGBoost (XGB), were trained on a part of the data. Their efficiency and accuracy in predicting experimentally identified ballistic limit features were then evaluated for the remaining data subset. By leveraging available data and incorporating specific physical mechanisms into machine-learning algorithms, the study analyses the applicability and limitations of these approaches.