Physics-Reinforced Deep Learning for Structural Metamodeling

这项研究将开发新的计算方法，以推进地震下土木结构系统的建模、分析和评估。目前在结构分析中广泛使用的简化或降阶模型在准确预测地震作用下结构反应的复杂非线性（例如存在较大的非线性漂移、损伤、破坏等）方面存在严重的局限性。人工智能（AI）在土木工程中的应用为结构建模提供了一种强大的新方法。然而，目前训练一个可靠的AI模型需要大量的数据，即使有丰富的数据，训练出来的模型也很难解释，没有什么物理意义。为了解决这些基本问题，并弥合人工智能与基于性能的工程之间的知识差距，该项目将整合深度学习和物理原理，以便在地震灾害下对非线性结构进行有效和概率建模。该研究将以更少的计算量促进更有效的工程结构设计、可靠性分析、控制和优化。该项目还将建立一个综合的研究-教育-推广计划，该计划将(i)转变对基于特定领域知识的机器学习的基本理解，（ii）促进本科生，特别是女性和少数民族的参与，以及（iii）激励高中生追求stem相关的职业。这项研究将机器学习的新元素引入到基于性能的结构工程中，通过将深度学习与物理知识相结合，为结构系统的地震反应建模开辟了一条新的途径。本项目的具体研究目的包括：(1)开发基于无监督学习的地震动选择，以优化地震反应数据库的生成；(2)为结构元模型的创新、物理强化的深度学习范式建立严格的公式和算法；(3)开发神经网络压缩方法，对元模型进行精简，以实现有效的推理；(4)将可变性纳入概率地震反应预测和脆弱性分析。由此产生的基于深度学习的元模型将具有显著的特征，包括(i)具有物理意义的可解释性，（ii）对未知情况的泛化和外推，（iii）基于轻量级/压缩元模型架构的实时推理，以及（iv）处理不完整和恐惧数据的能力。该方法将适用于各种复杂载荷条件下的非线性动力系统。该项目将推进非线性结构动力学、计算建模、鲁棒高效机器/深度学习、优化和不确定性量化等多学科的知识基础。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This research will develop new computational methods to advance modeling, analysis and assessment of civil structural systems subjected to earthquakes. Currently available simplified or reduced-order models widely used in structural analysis have severe limitations in accurately predicting the complex nonlinearities in structural response under earthquake loading (e.g., in the presence of large nonlinear drifts, damage, failure, etc.). The infusion of artificial intelligence (AI) into civil engineering offers a powerful new approach for structural modeling. However, currently large amounts of data are required to train a reliable AI model, and even with rich data, the trained models are difficult to interpret and have little physical meaning. To address these fundamental issues and to bridge the knowledge gap between AI and performance-based engineering, this project will integrate deep learning and physics principles for efficient and probabilistic modeling of nonlinear structures under earthquake hazards. This research will advance more efficient design, reliability analysis, control and optimization of engineering structures with much less computational efforts. The project will also establish an integrated research-education-outreach program that will (i) transform the fundamental understanding of machine learning grounded with domain-specific knowledge, (ii) promote participation by undergraduates, in particular, women and minorities, and (iii) inspire high school students to pursue STEM-related careers.This research will breathe novel elements of machine learning into performance-based structural engineering, opening a new avenue by leveraging deep learning integrated with physics knowledge for modeling of seismic response of structural systems. The specific research aims of this project include: (1) developing unsupervised learning-based ground motion selection for optimal generation of seismic response database, (2) establishing rigorous formulation and algorithm for an innovative, physics-reinforced deep Learning paradigm for structural metamodeling, (3) developing a neural network compression approach to prune the metamodels for efficient inference, and (4) incorporating variability for probabilistic seismic response prediction and fragility analysis. The resulting deep-learning-based metamodels will possess salient features including (i) interpretability with physical meaning, (ii) generalizability and extrapolation to unseen cases, (iii) real-time inference based on a lightweight/compressed metamodel architecture, and (iv) capability of addressing incomplete and scare data. This approach will be applicable to a wide range of nonlinear dynamical systems under complex loading conditions. This project will advance the knowledge base in multiple disciplines of nonlinear structural dynamics, computational modeling, robust and efficient machine/deep learning, optimization and uncertainty quantification.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.