Simulation-augmented learning for ballistic performance prediction of ultra-high-strength steel

Direct measurement of ballistic service performance in extreme-service materials is costly and data-limited because penetration testing requires specialized facilities and yields only a small number of valid observations. This limitation makes it difficult to establish reliable links between routine mechanical properties and terminal ballistic responses. To address this gap, we propose a simulation-augmented two-stage learning framework that maps standard mechanical-test results to service-level ballistic responses under small-sample conditions. The framework first identifies the Johnson-Cook (JC) constitutive and failure parameters, together with a pressure-cutoff criterion, from limited experiments using chaos-based global optimization. The identified parameters are then used to perform high-throughput penetration simulations, thereby augmenting the scarce experimental service data. The augmented dataset is processed using a two-stage learner, in which random forests classify discrete outcomes and a multilayer perceptron regression predicts continuous metrics. In a case study of G33 ultra-high-strength steel, the optimized feature-selected model achieved cross-validated prediction accuracies of 98.3%, 74.8%, 92.9%, and 89.8% for penetration occurrence, post-penetration projectile integrity, residual velocity, and critical projectile fragmentation velocity, respectively. The proposed framework provides a transferable strategy for ballistic performance prediction in other impact-dominated material systems with limited service data.