Simulation-Admissibility-Aware Surrogate Optimization of Electric-Vehicle Side-Pole Crashworthiness
Abstract
Abstract Crashworthiness surrogates are usually trained on successful finite element (FE) simulations, although automated full-vehicle workflows may include mixed-variable designs that fail during reconstruction, meshing, solution, or response extraction. This study proposes a simulation-admissibility-aware surrogate-assisted framework for electric-vehicle side-pole crashworthiness design. A battery-electric SUV was parameterized with 32 continuous geometry and 36 discrete material/thickness variables. Of 600 SFE–LS-DYNA designs, 528 were admissible. A workflow-admissibility classifier was evaluated by repeated stratified nested cross-validation. In the admissible domain, the structure-informed MS-GIN surrogate achieved a mean R2 of 0.91 and NRMSE of 6.5%. Bootstrap SHAP identified influential variables in the B-pillar and lower body. NSGA-II optimization with FE verification yielded a compromise design reducing body mass, B-pillar chest intrusion, and battery-neighborhood intrusion by 3.6%, 9.5%, and 14.6%, while satisfying the other constraints. The framework captures admissibility risk, structural sensitivity, and lightweighting–crashworthiness trade-offs in mixed-variable crash design.