Prediction of failure modes, strength, and deformation capacity of RC shear walls through machine learning

Abstract

Shear walls are typically the major lateral load-carrying components in high-rise buildings owing to their high lateral strength and stiffness. This study introduces a technique for predicting the seismic performance of reinforced concrete (RC) walls using machine learning (ML) methods. Based on various ML algorithms, predictive models were developed using a database containing experimental data of 429 RC walls collected from the literature. The performance of each predictive model was discussed in detail. The results indicated that the XGBoost and GB algorithms accurately predicted the failure modes of RC walls with an accuracy of 97%. The gradient boosting and random forest algorithms performed best in predicting the lateral strength and ultimate drift ratio of RC walls, with a mean predicted-to-tested strength ratio of 1.01 and a predicted-to-tested ultimate drift ratio of 1.08. The flexure-to-shear strength ratio and shear-to-span ratio of RC walls had a greater influence on the failure modes of RC walls, with a relative importance factor of 54.1% for these two characteristics. Boundary longitudinal reinforcement, shear-to-span ratio, and distributed reinforcement had an obvious influence on the lateral strength of RC walls with summation of relative important factors of over 80%. The greatest influence on the ultimate drift ratio of RC walls is shear-to-span ratio, with a relative importance factor of 34.1%. The comparisons indicate that the model developed in this study can predict the shear strength and flexural strength of RC walls more accurate and efficient than the existing design formulas. Furthermore, a brief boundary that could separate flexure-shear failure from other failure modes is provided based on the ML models. Finally, a graphical user interface (GUI) platform is created to facilitate the practical design of RC walls.