Risk-aware surrogate-assisted multi-objective design optimization of profiled steel-concrete composite slabs

Abstract

Profiled steel-concrete composite slabs require a balance among self-weight, service performance, and ultimate load resistance. Direct finite element (FE)-based optimization is computationally expensive because every candidate requires a nonlinear analysis. This study presents a risk-aware surrogate-assisted framework that incorporates calibrated prediction uncertainty directly into multi-objective optimization. An audited dataset of 5000 ANSYS design points was used to train XGBoost models for service deflection and steel-deck stress and heteroscedastic deep ensembles for ultimate capacity and concrete compressive strain. Calibration on held-out residuals produced uncertainty-adjusted screening bounds. NSGA-II used upper-bound predictions for service and strain constraints and a lower-bound prediction for the capacity objective. The service models achieved test \(R^2\) values of 0.974 for deflection and 0.970 for deck stress, while the ultimate-response models achieved values of 0.751 for capacity and 0.689 for concrete strain. Across 30 independent NSGA-II runs, the final normalized hypervolume was \(0.8940\pm 0.0032\), and 29 runs satisfied the adopted terminal-stability criterion. Mean-only optimization produced 66 of 200 candidates that failed the adopted surrogate-bound criteria after uncertainty-aware rechecking, whereas all 200 risk-aware candidates satisfied those criteria. An equal-weight closest-to-ideal rule selected solution RA-P088 as the representative compromise, with a mass of 723.4 kg, conservative deflection of 0.2394 mm, and conservative capacity of 326.48 kN. FE re-analysis of this candidate also satisfied the adopted screening limits. The framework supports rapid trade-off screening and targeted FE confirmation of shortlisted designs but does not replace project-specific structural verification.