Leveraging machine learning techniques to support a holistic performance-based seismic design

Abstract

The increasing vulnerability of communities to natural hazards motivates novel design and assessment methods to ensure that the built environment performs optimally during its lifetime. The current design methodologies do not account for life-cycle impacts across multiple performance domains such as economy and environment. Therefore, low-effort and designer-centric computational methods are needed to support a multi-objective performance-based design, from the conceptual stage to design development. This study presents a framework for a holistic performance-based seismic design of buildings. The proposed framework leverages machine learning techniques to extract the implicit, and highly complex, relationship between design parameters, geometric configuration, and performance measures. At early design, data-driven surrogate models (trained on performance inventories) are used to identify candidate structural systems and their approximate design parameters. At the detailed design stage, a deep learning-based engine generates seismic risk estimates based on simpler nonlinear static analysis on the candidate systems or their equivalent low-order dynamic models. A case study illustrates the framework’s application for performance-based seismic design of multistory commercial buildings in Charleston, SC.