A dimensional-analysis-guided two-stage surrogate for perforation screening of steel-concrete-steel composite walls under rigid projectile impact

Steel-concrete-steel (SC) composite walls are used in protective and nuclear structures, but perforation assessment under rigid projectile impact still depends on costly nonlinear simulation or conservative code procedures. This study develops a dimensional-analysis-guided two-stage surrogate for double-skin SC walls subjected to normal impact by rigid flat-nose projectiles. A calibrated and transfer-validated LS-DYNA template was used to generate a quality-controlled accepted FE database of 300 cases over an eight-dimensional design space, which was reduced to five dimensionless groups using Buckingham-Π analysis. Perforation was classified with a probability-calibrated Random Forest, and transmitted residual velocity was predicted conditionally using Gaussian process regression. Under nested 5-fold cross-validation, the cascade achieved ROC-AUC = 0.970 ± 0.020 and end-to-end RMSE = 17.1 ± 8.2 m/s; these metrics quantify FE emulation fidelity rather than direct experimental predictive accuracy. Mechanistically, the surrogate indicates that local perforation is governed primarily by the competition between concrete resistance energy and projectile kinetic energy, with steel-plate and concrete thickness acting as co-controlling ratios within the calibrated domain.