University of British Columbia
Presentation Title: Impacts of M9 Cascadia Subduction Zone Earthquakes on the Seismic Performance of Tall Non-Ductile Reinforced Concrete Shear Wall Buildings
Co-Author: Preetish Kakoty
Abstract: Over 300 tall buildings constructed prior to the advent of modern seismic design standards (i.e., 1960s to 1980s) can be found in the City of Vancouver’s West End neighborhood. These reinforced concrete shear wall structures are characterized by thin walls, of 8 inches or ~200 mm in thickness, with a single layer of reinforcement and a lack of boundary zones. These structures represent only 20% of buildings in the West End, but they house over 40% of the neighborhood’s population. While the local engineering community generally regards these structures as seismically vulnerable, their expected performance in future earthquakes is largely unknown. In this study, we quantify the (i) collapse risk, (ii) expected direct economic losses and (iii) recovery times of a series of 25 representative building archetypes, with heights ranging from 10 to 30 stories, under a M9 Cascadia subduction zone earthquake scenario. Structural response parameters are obtained from nonlinear response history analyses of three-dimensional structural analysis models conducted in OpenSeesPy. These results are translated to expected economic losses using the FEMA P-58 framework as implemented in the SimCenter’s Pelicun tool. In turn, those results are leveraged to evaluate the time to achieve two distinct recovery states: (a) shelter-in-place and (b) functional recovery, using the open-source Tool for Recovery Estimation and Downtime Simulations (TREADS). Under the M9 scenario considered, the collapse risk of the archetype buildings ranges from 10% to 40%. The mean loss ratio, i.e., repair cost normalized by the building replacement cost, ranges from 54% to 82%. The probability of the buildings having sheltering capacity immediately post-earthquake ranges from 52% to 93%. Lastly, the median time to achieve functional recovery ranges from 450 to 865 days. These results serve to quantify the elevated, and highly variable, seismic risk of this vulnerable taxonomy of buildings and highlight the need for risk mitigation strategies.
University of Michigan
Presentation Title: A Deep Learning-based Multi-Fidelity Monte Carlo (DL-MFMC) scheme for efficient reliability analysis of nonlinear structural systems subject to natural hazards
Co-Author: Liuyun Xu
Abstract: To maximize the benefits of performance-based engineering, a key challenge emerges in the efficient estimation of reliabilities associated with inelastic limit states under stochastic loads modeling the actions of natural hazards. Notwithstanding the potential of high-fidelity models for accurately capturing essential features of inelastic behavior, they quickly turn computationally infeasible when adopted within modern probabilistic assessment frameworks that require a considerable number of nonlinear analyses. To minimize the computational cost while preserving accuracy, a deep learning-based multi-fidelity Monte Carlo (DL-MFMC) scheme is presented for efficiently characterizing uncertainties and predicting inelastic responses and therefore estimating probabilities of failure. In this approach, all available high-fidelity numerical simulations are simultaneously utilized as the training data of a long short-term memory (LSTM) deep learning network, which then serves as a cheap low-fidelity model within the MFMC setting. Consequently, the correlation between the high- and low-fidelity models is enabled. Through an application to a high-rise wind-excited steel building, the proposed scheme is demonstrated to be capable of estimating reliabilities subject to stochastic wind excitation given various limit states of interest. Moreover, the DL-MFMC scheme is shown to be not only significantly faster than state-of-the-art high-fidelity probabilistic frameworks but also remarkably accurate in reproducing inelastic responses of structural models characterized by complex material behaviors (e.g., steel fatigue) and initial imperfections.
University of Notre Dame
Presentation Title: Automating assembly-based visual damage detection to accelerate learning from disasters
Co-Author: Rachel Hamburger
Abstract: Mounting losses due to hurricanes has created an acute crisis in the US. Devastating as they may be, each of these hurricanes holds within them valuable lessons that are central to reversing this alarming trend. Efforts to maximize this learning potential have spurred the systematic collection and dissemination of critical field observations. Access to these large-scale datasets, when coupled with rapid advancements in open-source AI, can transform our ability to efficiently estimate disaster impacts on the built environment. This study particularly takes advantage of the growth in computer vision-based techniques to automate performance assessments of buildings. Applications of this technology typically focus on one of two extremes: global damage assessment (typically through binary classifiers) or highly localized damage classification on individual components, leaving a vital gap relevant to hurricane catastrophe modeling, where global performance is assigned by aggregating damage of the most vulnerable subassemblies. In response, this study proposes a hierarchical workflow to virtually characterize wind-induced damage of the building stock at the assembly-level from rapid imaging data sources. Input images are preprocessed and filtered using a series of binary neural network classifiers. Retained photos are analyzed via a segmentation task, where subassembly information is calculated from visual damage indicators for each component class. This unique combination of classification and segmentation models is capable of achieving higher granularity damage detections consistent with the component-based damage classification methods standardized within the natural hazards research community and promoted by conventional loss models like HAZUS. Performance of the workflow is demonstrated using large reconnaissance datasets collected after major hurricanes to further demonstrate its utility in automating the damage quantification process as part of ongoing efforts to more swiftly learn from disaster.