A comparative assessment of mechanistic and data-driven models to estimate building responses

Abstract

The second-generation performance-based seismic design (PBSD) framework enables structural engineers to target specific stakeholder-drive building performance objectives. However, due to the labor-intensive and computationally processes of performing iterative response history analyses (NRHAs) and subsequent damage, loss, and downtime assessments, it has not been widely adopted in engineering practice. To address this challenge, several simplified methodologies have been developed for estimating building seismic response demands (i.e. so-called engineering demand parameters). One common theme among these models is that they are all rooted in the fundamental principles of structural dynamics and beam theory. However, while some rely solely on these basic physics and engineering principles, others have attempted to integrate basic statistical regression using structural response data generated from NRHAs performed on parametric structural models. In this study, the authors lay out a spectrum of methodologies for estimating structural response demands to be utilized in PBSD. On one end of the spectrum is a simplified purely mechanistic method, which is often preferred by practicing engineers because it is typically highly generalizable and easy to interpret. However, this method often relies on many simplifications which can influence the accuracy of the response estimates. On the other end of the spectrum is a purely data-driven model which utilizes parametric datasets generated from NRHAs, which is often viewed as a black box. However, these models are less reliant on the convenient simplifications that are adopted in the simplified mechanistic models and thus might achieve higher accuracy. Between these two extremes, there are models that combine elements of basic physics and statistical learning. This paper starts by introducing the development of a comprehensive database, which includes seismic designs of 621 steel moment resisting frames (SMRFs), the corresponding nonlinear structural models, and associated seismic responses (i.e., peak story drifts, peak floor accelerations, and residual story drifts). Then four existing methodologies that fall within the spectrum of seismic response estimation approaches are introduced, critically examined to reveal their benefits and drawbacks, systematically evaluated to quantitatively illustrate the accuracy. Inspired by these existing methods, one hybrid (mechanistic + data-driven) and one purely data-driven models are rigorously developed via training, testing, and validating against the database. Finally, a comparative assessment among mechanistic, hybrid, and data-driven models is performed.