Machine learning‐based resistance prediction model for compressive arch action in concrete beam–column substructures

Abstract With the rapid development and application of machine learning techniques, the use of proper intelligent algorithms to investigate the progressive collapse resistance of concrete frame structures has become significant. In this paper, a comprehensive database of 134 experimentally tested one‐way beam–column subassemblies was compiled to investigate the compressive arch action (CAA) capacity under a column removal scenario. Three models, including backpropagation neural network (BPNN), support vector regression (SVR), and gene expression programming (GEP), were developed and systematically compared in terms of prediction accuracy, generalization ability, model complexity, and interpretability, followed by the evaluation of each model's performance by various statistical criteria. It is found that the BPNN model provides reasonable accuracy, but its black‐box nature limits practical implementation. The SVR model, despite its strong fitting performance on training datasets, suffers from severe overfitting and shows low reliability for generalization. In contrast, the GEP model achieves comparable accuracy to BPNN, while offering explicit mathematical formulations and reduced computational complexity, making it a superior choice for practical applications. To further enhance predictive capability, a GEP‐based analytical model is formulated and validated against existing analytical models, demonstrating exceptional efficiency and accuracy in CAA capacity prediction. Parametric analyses confirm that the GEP‐based model reliably captures fundamental engineering relationships, reinforcing its suitability for progressive collapse assessment. Given its robustness and computational efficiency, the GEP‐based model is recommended as a practical tool for evaluating CAA capacity in reinforced concrete structures.