MCA: Using Machine Learning to Predict Seismic Failure Limit States in Buildings

这项中期职业发展(MCA)奖将使首席调查员(PI)能够接受机器学习(ML)方法的培训，以弥合基于性能的地震工程和ML之间的差距，以估计结构的倒塌极限状态。结构倒塌是建筑物最具灾难性的破坏模式，也是最难用传统方法进行评估的。本研究将应用最大似然算法来预测建筑物在强震作用下的结构倒塌。这些算法将在没有明确编程的情况下接受学习训练，并可以改变设计和评估结构系统的方式。基于性能的方法被用于重要结构的评估或设计，但由于模型的复杂性以及漂移和其他响应参数对小的输入参数变化的敏感性，估计结构失效的方法有很大的局限性。此外，大多数建筑都是用简化的弹性方法设计的，在自然灾害事件下的预期结构破坏只能从一般的考虑因素中近似计算出来。在这项研究中，ML算法将使用数值模拟和实验测试结果进行训练，以有效地预测结构设计方案的倒塌。与斯坦福大学和麻省理工学院的研究伙伴的协同合作将建立PI在这一领域的研究能力。这项研究将得到一个基于高中外展、对研究生和本科生的支持以及培训示范的教育计划的补充。该奖项将为国家科学基金会(NSF)在国家减少地震灾害计划中的作用做出贡献。项目数据将存档在美国国家科学基金会支持的自然灾害工程研究基础设施(NHERI)数据仓库(https://www.designsafe-ci.org).目前的崩溃方法基于复杂的非线性有限元模型，这在设计过程中可能是一项繁重的任务，甚至对于现有系统的性能评估也是如此。该项目的研究目标包括：(I)应用于失效极限状态的最优ML技术的实施；(Ii)从现有数据库中挖掘建筑结构部件的动态响应的数据挖掘；以及(Iii)改进结构系统性能的方法的发展。将考虑几种有前景的结构倒塌应用策略，如人工神经网络、支持向量机和响应面模型的变体。将回答以下基本问题：(I)需要什么级别的建筑输入数据才能有效地预测坍塌？以及(Ii)是否可以训练ML算法来评估受损系统的储备能力和冗余度，在受损系统中，材料劣化是高度不确定的，或者关键结构部件(例如柱子)被移除？ML算法将被用来发现与结构倒塌相关的隐藏关联，并将被训练来考虑几组输入数据，从基本建筑信息到非线性恶化的输入参数。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

This Mid-Career Advancement (MCA) award will enable the Principal Investigator (PI) to train in machine learning (ML) methodologies in order to bridge the gap between performance-based earthquake engineering and ML to estimate the collapse limit state of structures. Structural collapse is a building’s most catastrophic failure mode and the most difficult to evaluate using traditional methodologies. This study will apply ML algorithms to predict structural collapse of buildings under strong seismic events. These algorithms will be trained to learn without being explicitly programmed and can transform the way in which structural systems are designed and evaluated. Performance-based methodologies are used for evaluation or design of important structures, but there are significant limitations to estimate structural failure due to model complexity and the sensitivity of drift and other response parameters to small input parameter variations. Also, most buildings are designed using simplified elastic methods, and the expected structural damage under natural hazard events is only approximated from general considerations. In this study, ML algorithms will be trained using numerical simulations and experimental test results to efficiently predict collapse of structural design alternatives. The synergistic collaboration with research partners at Stanford University and the Massachusetts Institute of Technology will build the PI's research capabilities in this area. The study will be complemented by an educational program based on high school outreach, support of graduate and undergraduate research students, and training demonstrations. This award will contribute to the National Science Foundation (NSF) role in the National Earthquake Hazards Reduction Program. Project data will be archived in the NSF-supported Natural Hazards Engineering Research Infrastructure (NHERI) Data Depot (https://www.designsafe-ci.org). Current collapse methodologies are based on sophisticated nonlinear finite element models, which can be an onerous task in the design process and even for performance evaluation of existing systems. The project research objectives include: (i) implementation of optimal ML techniques for application to failure limit states, (ii) data mining of dynamic response of building structural components from available databases, and (iii) development of approaches to improve the performance of structural systems. Several promising strategies for the researched structural collapse application will be considered, such as variations of artificial neural networks, support vector machines, and response surface models. The following fundamental questions will be answered: (i) what level of building input data is required to efficiently predict collapse? and (ii) can ML algorithms be trained to assess the reserve capacity and redundancy of damaged systems, in which material deterioration is highly uncertain or key structural components (e.g., columns) are removed? The ML algorithms will be used to find hidden correlations associated to structural collapse and will be trained to consider several sets of input data, ranging from basic building information to nonlinear deteriorating input parameters.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.