Artificial neural network surrogate modeling for explosively formed projectiles using space filling sampling

This study provides a surrogate framework that combines high fidelity Finite Element Method ( FEM ) calculations and Artificial Neural Network ( ANN ) topologies, to efficiently evaluate the multiphysical performance of Explosively Formed Projectiles ( EFP ). The geometric domain was discretized with 64 design samples using a space filling approach. Physically unstable configurations were eliminated, resulting in a refined dataset of 62 samples, demonstrating high data efficiency compared to exhaustive grid-based sampling strategies. The regularized neural network was compared to Response Surface Methodology, Support Vector Regression and Random Forest models using a Repeated K Fold Cross Validation methodology. Mathematical solutions of severe nonlinear mechanical dependencies were obtained by optimization using the Limited memory Broyden Fletcher Goldfarb Shanno ( L-BFGS ) algorithm combined with L 2 regularization. Standard machine learning models overfit the chaotic physical phases such projectile stretching, as demonstrated by rigorous cross validation, while the neural network exhibited superior out-of-sample generalization for kinematic and morphological parameters, and remained highly competitive for terminal ballistic variables. The kinematic velocity was statistically evaluated with a mean determination coefficient of 0.9977 and penetration depth of 0.8091. Live-fire tests validated the framework with relative errors below 6% for velocity and 15% for crater growth. Rather than a standalone inverse-design tool, this framework provides a data-efficient rapid screening methodology for precision-guided munitions, significantly reducing computational costs prior to empirical verification.