Probabilistic machine leaning models for predicting the maximum displacements of concrete-filled steel tubular columns subjected to lateral impact loading

• Published in: Engineering applications of artificial intelligence Vol. 135; p. 108704
• Main Authors: Lai, Dade, Demartino, Cristoforo, Xiao, Yan
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.09.2024
• Subjects: CFST columns; Lateral impact; Machine learning; Maximum displacements; NGBoost; Probabilistic model; SHapley additive exPlanation (SHAP); Machine learning; SHapley additive exPlanation (SHAP); Probabilistic model; Maximum displacements; CFST columns; NGBoost; Lateral impact
• ISSN: 0952-1976; 1873-6769
• DOI: 10.1016/j.engappai.2024.108704

The evaluation of the maximum displacement is pivotal for the application of the performance design of Concrete Filled Steel Tubes (CFSTs) under lateral impact loads. Traditional approaches have limitations in their predictive capabilities and necessitate substantial modeling efforts and computational resources, especially if employed for probabilistic predictions. In recent years, Machine Learning (ML) algorithms have been increasingly utilized to tackle complex problems involving extreme loads like impact and blast. However, a significant drawback of most ML models is their limitation in accounting for uncertainties in their outputs. This study endeavors to introduce a novel probabilistic ML model utilizing the algorithm known as NGBoost, which leverages gradient boosting to enable generic probabilistic predictions. The predictions obtained from NGBoost, using both Normal and LogNormal distributions, are compared with other deterministic models such as eXtreme Gradient Boosting (XGBoost), artificial neural network (ANN), and with a different probabilistic model, i.e., Gaussian Process Regression (GPR). In order to train and validate the models, a comprehensive database was compiled, consisting of 192 experimentally tested specimens. The results demonstrate that the probabilistic ML model achieves high accuracy and provides more detailed probabilistic predictions. Furthermore, the SHapley Additive exPlanations (SHAP) method is utilized to evaluate the relative importance of the input features and establish the relationship between the input features and the target output. Additionally, a comprehensive parametric study is conducted to explore the influence of each input feature. Finally, a simple application of the probabilistic model is presented. The proposed probabilistic ML model presents an application in performance-based design that boasts decreased computational demands and enhanced user-friendliness compared to conventional approaches. •Database of 192 square/circular CFST columns tested under lateral impact loads.•Four ML regression algorithms predict maximum displacements.•NGBoost model achieves high accuracy and probabilistic predictions.•SHAP algorithm assesses input features’ significance.•Parametric study investigates influential input features.