A systematic investigation of ANN structures for predicting optimal dimensions of reinforced concrete retaining walls on strong soil

Reinforced concrete retaining walls (RCRWs) are among the most widely used earth-retaining structures in civil engineering. Traditional preliminary design of RCRWs relies on experience-based assumptions and iterative procedures, which may result in suboptimal dimensions and inefficient material utilization. Existing artificial intelligence (AI)-based research predominantly focuses on safety factor estimation, but accurately estimating crosssectional dimensions is more critical to achieve an efficient design. This study presents a novel data-driven framework combining differential evolution algorithm (DEA) with artificial neural network (ANN) model evaluation for predicting optimal cross-sectional dimensions of RCRWs on strong soils. Unlike previous studies, this research systematically explores four distinct ANN model structures and examines the effects of the variation in the selected output parameters on the performance of the models. A comprehensive dataset comprising 175 optimally designed RCRWs was generated via DEA, considering wall height, surcharge load, and backfill internal friction angle as key variables. Systematic model comparisons revealed that Model-1 achieved the best performance (R2 = 0.9997, MaxAE = 0.0285 m). Additionally, SHAP (SHapley Additive exPlanations) analysis was conducted to interpret the decision-making process of the model, confirming that the predictions are physically consistent with geotechnical design principles. Comparative analyses with traditional and regression-based approaches revealed that the ANN model consistently outperforms existing methods in both accuracy and generalization capability. This work provides near-optimal initial dimension estimates by eliminating traditional iterative design inefficiencies. The results show that intelligent preliminary design can achieve professional-level accuracy and that it is more effective to shift from traditional approaches to AI-based approaches in geotechnical design applications.