Strength prediction of ECC-CES columns under eccentric compression using adaptive sampling and ML techniques.

A novel type of concrete-encased steel (CES) composite column implementing Engineered Cementitious Composites (ECC) confinement (ECC-CES) has recently been introduced, offering significantly enhanced failure behavior, ductility, and toughness when compared to conventional CES columns. This study presents an innovative method for predicting the eccentric compressive capacity of ECC-CES columns, utilizing adaptive sampling and machine learning (ML) techniques. Initially, the research introduces a finite element (FE) model for ECC-CES columns, incorporating material and geometric nonlinearities to capture the inelastic behavior of both ECC and steel through appropriate constitutive material laws. The FE model was validated against experimental data, demonstrating strong predictive accuracy. An adaptive sampling process was employed to efficiently explore the design space, resulting in a database of 2,908 FE models. Subsequently, six machine learning models were used to predict the eccentric compressive capacity based on the generated FE database. These models were thoroughly evaluated and demonstrated superior prediction accuracy compared to established design standards like EC4 and AISC360. Based on evaluation metrics, the Gaussian Process Regression (GPR), CatBoost (CATB), and LightGBM (LGBM) models emerged as the most accurate and reliable, with over 97% of the finite element (FE) samples falling within a 10% error range. While the ML models demonstrate impressive performance, their black-box nature restricts their practical use in design applications. Consequently, this study introduces a proposed design that offers competitive performance metrics. The novelty of this work lies in integrating adaptive sampling through Bayesian Optimization (BO) with the power of machine learning (ML) to generate training data that effectively covers a large input space while minimizing error. SVR, CatBoost, and GPR models demonstrated mean μ, R, and a20-index values near 1.0, with CoV and MAPE% values consistently low, indicating highly accurate predictions across testing subsets.