Response-preserving surrogate acceleration of SHPB explicit-dynamics simulations: A mechanics-informed benchmark with event-centered supervision

Split Hopkinson pressure bar (SHPB) explicit-dynamics simulations provide high-fidelity forward maps from constitutive descriptors and loading conditions to bar-strain histories and reconstructed specimen response, but repeated evaluations remain computationally expensive. This work develops a response-preserving surrogate that predicts incident, reflected, and transmitted bar-strain histories rather than stress–strain curves directly, and then recovers specimen response through fixed SHPB reconstruction and event extraction. The simulation-derived benchmark contains 100,000 LS-DYNA samples over a representative metallic property domain, combined with stratified impact loading, Johnson–Cook admissibility checks, finite-pulse and loadability screening, and response labeling. Six sequence-aware temporal backbones are evaluated under multiple split protocols for waveform-history prediction. The results show that event-centered supervision improves reconstructed-response fidelity and failure-event localization, and that waveform-history prediction better preserves event-sensitive response than direct stress–strain prediction or non-temporal reduced-basis baselines. After training, the surrogate provides a reusable evaluator for repeated SHPB parameter queries while retaining the standard wave-to-response reconstruction chain. The framework therefore supports efficient constitutive-parameter exploration without replacing the mechanics-based post-processing protocol that defines the measured response.