Machine Learning-Based Framework for Failure Forecast and Shear Strength Estimation of Non-Conforming RC Shear Walls

ABSTRACT

Earthquake reconnaissance after the recent Kahramanmaras, Turkey, earthquake sequence in February 6, 2023 has shown that the majority of the reinforced concrete (RC) structural elements do not comply with the Turkish Building Earthquake Code (TBEC 2018) requirements. Accurate estimation of the actual strength of existing non-compliant RC structural members is critical to accomplish reliable earthquake performance assessments. This study aims to propose a machine learning-based framework to establish a realistic failure forecast as well as an accurate shear strength estimation depending on the expected failure modes of such walls. To achieve this, shear wall design properties are assigned as inputs whereas failure mode and shear strength were designated as outputs for the corresponding classification and regression problems, respectively. Widely used machine learning methods, namely: Ridge Linear Regression, Logistic Regression, Multi-layer Perceptron, Support Vector Machine, Random Forest, XGBoost, LightGBM, and CatBoost are employed using a database consisting of 432 non-conforming shear walls. The data is randomly split to train and test sets (80% and 20%, respectively) to develop the predictive models and to evaluate model performances, respectively. Performance metrics (e.g. $R^2$, RMSE) are evaluated over a hundred random splits to ensure robustness. The proposed CatBoost-based models achieve overall accuracies above 90% for both failure prediction and shear strength estimation. The proposed framework is believed to be a promising tool for earthquake engineers for practical and accurate failure mode forecast and shear strength estimation of non-compliant RC shear walls.

KEYWORDS:

• Non-conforming shear walls
• failure modes
• shear strength
• machine learning
• failure forecast

Disclosure Statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

The project has been supported by funds from the Scientific and Technological Research Council of Turkey (TUBITAK) under Project No: 218M535. Opinions, findings, and conclusions in this paper are those of the authors and do not necessarily represent those of the funding agency.