No more black-boxes: estimate deformation capacity of non-ductile RC shear walls based on generalized additive models

Abstract Machine learning techniques have gained attention in earthquake engineering for their accurate predictions, but their opaque black-box models create ambiguity in the decision-making process due to inherent complexity. To address this issue, numerous methods have been developed in the literature that attempt to elucidate and interpret black-box machine learning methods. However, many of these methods evaluate the decision-making processes of the relevant machine learning techniques based on their own criteria, leading to varying results across different approaches. Therefore, the critical significance of developing transparent and interpretable models, rather than describing black-box models, becomes particularly evident in fields such as earthquake engineering, where the interpretation of the physical implications of the problem holds paramount importance. Motivated by these considerations, this study aims to advance the field by developing a novel methodological approach that prioritizes transparency and interpretability in estimating the deformation capacity of non-ductile reinforced concrete shear walls based on an additive meta-model representation. Specifically, this model will leverage engineering knowledge to accurately predict the deformation capacity, utilizing a comprehensive dataset collected from various locations globally. Furthermore, the integration of uncertainty analysis within the proposed methodology facilitates a comprehensive investigation into the influence of individual shear wall variables and their interactions on deformation capacity, thereby enabling a detailed understanding of the relationship dynamics. The proposed model stands out by aligning with scientific knowledge, practicality, and interpretability without compromising its high level of accuracy.