Shen Feng, Jha Ishan, Isleem Haytham F, Almoghayer Walaa J K, Khishe Mohammad, Elshaarawy Mohamed Kamel
College of Civil Engineering, Huaqiao University, Xiamen, 361021, China.
Xiamen Metro Construction Co., LTD, Xiamen, 361001, China.
Sci Rep. 2025 Feb 20;15(1):6194. doi: 10.1038/s41598-025-90648-2.
Confined columns, such as round-ended concrete-filled steel tubular (CFST) columns, are integral to modern infrastructure due to their high load-bearing capacity and structural efficiency. The primary objective of this study is to develop accurate, data-driven approaches for predicting the axial load-carrying capacity (P) of these columns and to benchmark their performance against existing analytical solutions. Using an extensive dataset of 200 CFST stub column tests, this research evaluates three machine learning (ML) models - LightGBM, XGBoost, and CatBoost - and three deep learning (DL) models - Deep Neural Network (DNN), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM). Key input features include concrete strength, column length, cross-sectional dimensions, steel tube thickness, and yield strength, which were analysed to uncover underlying relationships. The results indicate that CatBoost delivers the highest predictive accuracy, achieving an RMSE of 396.50 kN and an R of 0.932, surpassing XGBoost (RMSE: 449.57 kN, R: 0.906) and LightGBM (RMSE: 449.57 kN, R2: 0.916). Deep learning models were less effective, with the DNN attaining an RMSE of 496.19 kN and R of 0.958, while the LSTM underperformed substantially (RMSE: 2010.46 kN, R: 0.891). SHapley Additive exPlanations (SHAP) identified cross-sectional width as the most critical feature, contributing positively to capacity, and column length as a significant negative influencer. A user-friendly, Python-based interface was also developed, enabling real-time predictions for practical engineering applications. Comparison with 10 analytical models demonstrates that these traditional methods, though deterministic, struggle to capture the nonlinear interactions inherent in CFST columns, thus yielding lower accuracy and higher variability. In contrast, the data-driven models presented here offer robust, adaptable, and interpretable solutions, underscoring their potential to transform design and analysis practices for CFST columns, ultimately fostering safer and more efficient structural systems.
约束柱，如圆端钢管混凝土（CFST）柱，因其高承载能力和结构效率而成为现代基础设施不可或缺的一部分。本研究的主要目的是开发准确的数据驱动方法来预测这些柱的轴向承载能力（P），并将其性能与现有的解析解进行对比。本研究使用了包含200个CFST短柱试验的广泛数据集，评估了三种机器学习（ML）模型——LightGBM、XGBoost和CatBoost，以及三种深度学习（DL）模型——深度神经网络（DNN）、卷积神经网络（CNN）和长短期记忆网络（LSTM）。关键输入特征包括混凝土强度、柱长、横截面尺寸、钢管厚度和屈服强度，并对这些特征进行了分析以揭示潜在关系。结果表明，CatBoost具有最高的预测精度，均方根误差（RMSE）为396.50 kN，相关系数（R）为0.932，超过了XGBoost（RMSE：449.57 kN，R：0.906）和LightGBM（RMSE：449.57 kN，R²：0.916）。深度学习模型的效果较差，DNN的RMSE为496.19 kN，R为0.958，而LSTM的表现则大幅落后（RMSE：2010.46 kN，R：0.891）。SHapley值法（SHAP）确定横截面宽度是最关键的特征，对承载能力有正向贡献，而柱长则是显著的负向影响因素。还开发了一个基于Python的用户友好界面，可实现实际工程应用的实时预测。与10种解析模型的对比表明，这些传统方法虽然具有确定性，但难以捕捉CFST柱中固有的非线性相互作用，因此精度较低且变异性较高。相比之下，本文提出的数据驱动模型提供了强大、适应性强且可解释的解决方案，凸显了它们在改变CFST柱设计和分析实践方面的潜力，最终促进更安全、更高效的结构体系的发展。
