Reliability-based quantification of the benefits of machine learning predictive models in seismic structural design and performance assessment

Machine Learning (ML) techniques have been used extensively in research within the field of structural engineering due to their high level of accuracy in predicting the behavior of different structural elements. In fact, the superior predictive performance relative to traditional statistical models is often suggested as the primary motivation for the adoption of ML models. However, the implications of such improvements in predictive accuracy in the design and performance of structural systems have not been studied. This paper presents a reliability-based investigation of the tangible benefits provided by ML models in terms of structural design and performance. To quantify these benefits, the increase in predictive accuracy is interpreted as a reduction in epistemic uncertainty. The specific focus is on a predictive model that estimates the drift capacity of reinforced concrete shear walls (RCSWs) with special boundary elements. The accuracy of an extreme gradient boosting (XGBoost) model relative to a basic linear regression equation is quantified in terms of reduced epistemic uncertainty. Then, using 36 RCSW archetype buildings, a Monte Carlo-based procedure is implemented to evaluate the implication of the improved predictive accuracy to seismic design and performance. The study provides insights into how much improvement in accuracy (i.e., ML relative to traditional model) is needed to have a tangible effect on the seismic design and performance.