18 - Deep learning and data analytics for assessing seismic performance of civil infrastructures

Abstract

This chapter introduces comprehensive studies on developing advanced deep learning approaches and data analytic techniques for assessing structural performance of civil infrastructures under natural hazards. Classical methods such as physics-based methods for nonlinear structural time history analysis typically require excessive computational efforts, especially when numerous simulations are required to account for stochastic uncertainties of external loads (e.g., Monte Carlo simulations or incremental dynamic analysis for fragility analysis). To address this issue, we present multiple deep learning frameworks for metamodeling of civil infrastructures for safety assessment. Three deep architectures, including a long short-term memory (LSTM) network, a physics-guided convolutional neural network (PhyCNN), and a physics-informed multi-LSTM network (PhyLSTM), are proposed for metamodeling of nonlinear structures against earthquakes. In particular, the physics-informed deep learning paradigm outperforms classical non–physics-guided data-driven neural networks. The performance of the proposed approaches is successfully demonstrated through four case studies including both numerical examples and field validation. The results show that the developed architectures are promising, reliable, and computationally efficient approaches for assessing seismic performance of civil infrastructures, and offers significant potential in seismic fragility analysis of buildings for reliability assessment.