Universal boosting ML approaches to predict the ultimate load capacity of CFST columns

Abstract

Establishing a universal machine learning (ML) model in structural engineering is vital for understanding how various parameters, like geometry and material properties, influence a structure's behavior. This study aims to create a comprehensive ML model that considers the impact of different cross-sectional parameters on the ultimate load capacity (ULC) of concrete-filled steel tube (CFST) columns. This model assists engineers in making informed design decisions. The study employs a large dataset of 3094 data points with diverse geometric and material properties of CFST columns. After adjusting input features, robust boosting ML models (Catboost, LightGBM, and XGB) are meticulously fine-tuned using grid search and fivefold cross-validation. Monte Carlo simulation is used for further assessment. The results demonstrate that the most accurate XGB model delivers impressive accuracy, comparable to or better than existing literature models that focused on a single CFST column cross-section. The chosen XGB model is then utilized for feature importance analysis, local performance assessment, and sensitivity analysis through 1-D and 2-D partial dependence plots. These analyses help assess the input's contribution and effect on ULC prediction for CFST columns.