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

Area of Science:

• Civil Engineering
• Materials Science
• Computational Mechanics

Background:

• Conventional concrete-encased steel (CES) columns have limitations in ductility and toughness.
• Engineered Cementitious Composites (ECC) offer enhanced material properties for structural applications.
• ECC-confined CES (ECC-CES) columns present a novel composite with superior performance characteristics.

Purpose of the Study:

• To develop an innovative method for predicting the eccentric compressive capacity of ECC-CES columns.
• To utilize adaptive sampling and machine learning (ML) for accurate capacity prediction.
• To compare ML model performance against established design codes.

Main Methods:

• Development and validation of a finite element (FE) model for ECC-CES columns, including material and geometric nonlinearities.
• Application of adaptive sampling via Bayesian Optimization (BO) to generate a comprehensive FE database (2,908 models).
• Training and evaluation of six ML models (GPR, CatBoost, LGBM, etc.) to predict eccentric compressive capacity.

Main Results:

• FE model validation demonstrated strong predictive accuracy against experimental data.
• ML models achieved high prediction accuracy, with GPR, CatBoost, and LGBM showing over 97% of samples within a 10% error range.
• ML models significantly outperformed EC4 and AISC360 design standards in predicting eccentric compressive capacity.

Conclusions:

• Machine learning, particularly GPR, CatBoost, and LGBM, offers a highly accurate method for predicting ECC-CES column capacity.
• The integration of adaptive sampling with ML effectively generates robust training data for complex structural analysis.
• While ML models excel in accuracy, their interpretability requires further research for direct design application, prompting the proposal of a new design approach.