Scalable Assessment of Urban Earthquake Resilience: A Novel Model-informed Deep Learning Paradigm

复原力是指一个系统有效吸收压力而不严重破坏其功能的能力。如果社区通过减灾和灾前准备，实现了维持重要社区功能的适应能力，并在灾害发生后迅速恢复，则社区具有抗震能力。社区的复原力取决于建筑群（一组具有共同功能的建筑物，如住房）和配套基础设施系统的表现。需要对建筑群的复原力进行可靠的评估，以量化地震后的城市功能和恢复情况。评估地震恢复力的一个基本挑战是如何可靠和有效地估计建筑物的损坏/故障概率，并具有针对给定地震严重度的城市水平的可扩展性。为了应对这一挑战并弥合现有的知识差距，这个灾难恢复力研究赠款（DRRG）项目将通过开发一个全新的、可扩展的人工智能模型来评估大型城市建筑群的地震恢复力。研究结果可用于大规模城市建筑群的抗震性能评价。这种模式有助于减轻地震风险的决策，为灾后响应和恢复规划提供信息，并有助于改善未来的建筑设计。此外，还将与一名高中物理教师合作，在课堂实施后，编写一个关于“地震复原力对我的社区意味着什么”的微型单元。为了使用记录的地面运动（GM）可靠地估计地震情景下城市地区大量建筑物的地震需求，需要：（a）切合实际地估计转基因严重程度（本项目中的谱加速度）不同时期、不同位置和不同土质的建筑物的变化;（B）GM时程的适当和最佳选择;以及（c）用于数值模拟的建筑物的详细结构模型。然而，它是昂贵的，进行详细的建模大量的建筑物，并进行非线性时程分析，这样的模型下，一个大的GM概率代表不同的情况。我们的项目将解决这个基本问题，并通过开发和实施一系列新颖的AI授权算法和方法来实现可扩展性。该项目的总体目标是通过开发一个全新的、可扩展的基于模型的深度学习框架，对大型城市建筑群的地震恢复力进行评估。这将通过以下方式实现：（1）开发贝叶斯深度学习方法，以模拟地震场景下不同振动周期、位置和土壤类别的谱加速度变化;（2）创建一种新的无监督深度自动编码器-分类器算法，用于地面运动聚类和选择;（3）建立一种创新的模型通知符号深度学习方法，用于详细的非线性结构模型的元建模;（4）确定代表性建筑模型的元模型使能的脆弱性函数，以评估地震恢复力，考虑多个失效标准和多个性能目标;以及（5）展示了旧金山弗朗西斯科湾区代表性建筑的研究框架。该奖项反映了NSF的法定使命，并被认为值得通过使用基金会的智力价值和更广泛的评估来支持。影响审查标准。

Resilience refers to a system’s ability to efficiently absorb stresses without significant disruption to its functioning. Communities are earthquake resilient if by mitigation and pre-disaster preparation, they achieve the adaptive capacity for maintaining important community functions and recover rapidly following disasters. Community resilience depends on the performance of building clusters (a set of buildings that serve a common function such as housing) and the supporting infrastructure systems. Reliable assessment of the resilience of building clusters is required for quantifying urban functionality and recovery after an earthquake. A fundamental challenge in evaluating earthquake resilience is how to reliably and efficiently estimate probability of damage/failure of buildings with scalability to the urban level for a given earthquake severity. To address this challenge and bridge the existing knowledge gap, this Disaster Resilience Research Grants (DRRG) project will enable assessment of earthquake resilience for large-scale urban building clusters via developing a fundamentally novel and scalable AI-empowered model. The outcome of this project can be used to evaluate earthquake resilience of building clusters in large-scale urban areas. This paradigm facilitates decision making for seismic risk mitigation, informs planning for post disaster response and recovery, and helps improve future building design. Furthermore, in collaboration with a high school physics teacher and following classroom implementation, a mini-unit will be developed on “What does earthquake resilience mean to my community?” which will be available to other science teachers across the country.In order to reliably estimate the seismic demand on a large number of buildings in an urban area under earthquake scenarios, using recorded ground motions (GMs), there is a need for: (a) realistically estimating a GM severity measure (spectral acceleration in this project) variation for buildings with different periods, at different locations and on different soil classes; (b) proper and optimal selection of GM time-histories; and (c) a detailed structural model of buildings for use in numerical simulations. However, it is prohibitively expensive to conduct detailed modeling of a large number of buildings and carrying out nonlinear time history analyses of such models under a large set of GMs probabilistically representing different scenarios. Our project will address this fundamental issue and allow for scalability through the development and implementation of a series of novel AI-empowered algorithms and methods. The overarching goal of this project is to enable the assessment of earthquake resilience for large-scale urban building clusters, through developing a fundamentally novel and scalable model-informed deep learning framework. This will be achieved by: (1) developing a Bayesian deep learning approach to model the variation of spectral acceleration over different periods of vibration, locations, and soil classes, given an earthquake scenario; (2) creating a new unsupervised deep autoencoder-classifier algorithm for ground motion clustering and selection; (3) establishing an innovative model-informed symbolic deep learning method for metamodeling of detailed nonlinear structural models; (4) determining metamodel-enabled fragility functions for representative building models to assess earthquake resilience, accounting for multiple failure criteria and multiple performance objectives; and (5) demonstrating the researched framework for representative buildings in the San Francisco Bay Area.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.