Shear capacity prediction of squat flanged walls using machine learning approach with CatBoost and CTGAN

Abstract. Reinforced concrete (RC) walls serve an important role in providing lateral strength and stiffness for structures against seismic forces and wind loads. Squat walls, defined as having a shear span ratio less than or equal to two, are commonly used in building construction. Among squat wall designs, flanged walls have gained popularity due to architectural benefits. However, the seismic behavior of flanged RC squat walls is not fully understood due to complexities introduced by their cross-sectional shape and varying shear span ratios. Current structural design codes aim to capture wall behavior but have limitations in accurately modeling flanged squat walls. Previous experimental work has enhanced comprehension, but key aspects remain unsolved, such as precise prediction of stiffness and effective flange length. Data-driven machine learning models show promise for predicting engineering properties of composite materials like concrete. However, limited publicly available data hinders developing highly predictive computational models. This study addresses prior data limitations by using conditional tabular generative adversarial network (CTGAN) to synthesize additional tabular data. CTGAN, an adversarial learning method for tabular datasets, can generate realistic synthetic rows conditioned on discrete features. Here, CTGAN was trained on a 152-specimen experimental database to synthesize 5,000 data points. The synthesized data were then used to train a CatBoost gradient boosting algorithm for predicting shear capacity of flanged squat walls. The model achieved a high R-squared value of 0.878 when tested on 152 unseen experimental specimens, demonstrating accurate prediction of shear capacity in real data.