Machine learning-based prediction and optimization of corrugated steel-reinforced concrete composite arches under impact loading

Corrugated steel-reinforced concrete composite (CSC) arches, as innovative structures commonly constructed serving as protective structures and shed tunnels, are highly prone to suffering impact loading from falling rocks and heavy object collisions, threatening the structural stability and safety. To precisely assess the structural dynamic response of CSC arches, a data-driven approach of a machine learning-based neural network was developed. A comprehensive database comprising 300 numerical results was initially constructed via Latin Hypercube Sampling (LHS), and a modular neural network was proposed for predicting maximum deformations and classifying failure modes of CSC arches, in which active learning strategies and an oversampling technique are introduced to address the issues of dataset scarcity and class imbalance. The accuracy of the proposed neural network was verified against the true results on the test set; the comparison results revealed that the proposed model achieved an outstanding accuracy of R²= 0.9863 for the prediction task and a high accuracy of AUC = 0.9860 for the classification task, significantly outperforming the classical machine learning models (XGBoost, AdaBoost, LightGBM, RF, et al.). Following this, the Shapley additive explanations were employed to quantify the influence of geometric dimensions, material strength and impact parameters on the structural dynamic response, where the impact velocity, concrete thickness, and steel ratio were identified as the dominant parameters. Finally, a hybrid optimization strategy combining the Non-dominated Sorting Genetic Algorithm II (NSGA-II) with Multi-Objective Bayesian Optimization (MOBO) was proposed to perform multi-objective optimization. It yielded identify Pareto-optimal solutions that balance maximum deformation, self-weight, and material cost.