This research presents a novel methodology that uses Temporal Convolutional Networks (TCNs), a state-of-the-art deep learning architecture, for predicting the time history of structural responses to seismic events. By leveraging accelerometer data from instrumented buildings, the proposed approach complements traditional structural analysis models, offering a computationally efficient alternative to nonlinear time history analysis. The methodology is validated across a broad spectrum of structural scenarios, including buildings with pronounced higher-mode effects and those exhibiting both linear and nonlinear dynamic behaviors. Applications demonstrate high prediction accuracy across diverse building types, using datasets from the California Strong Motion Instrumentation Program (CSMIP). Training dataset development is grounded in core principles of structural dynamics, with results interpreted through the lens of earthquake engineering. Additionally, a pilot study is presented to reconstruct ground motions using the measured responses of a building. Despite recognizing limitations such as dataset size and model generalizability, the study highlights the transformative potential of advanced Artificial Intelligence (AI) techniques in seismic response prediction. Future research will investigate complementary strategies, including physics-informed AI, transformer architectures, and neural operators, to further enhance prediction accuracy. These advancements pave the way for improved Structural Health Monitoring (SHM) and support the evolution of Performance-Based Earthquake Engineering (PBEE) methodologies.
Two-page summary: click here.
Download full report: click here.
Full List of PEER Reports: click here.
