Deep learning enabled rapid nonlinear time history wind performance assessment

• Published in: Structures (Oxford) Vol. 66; p. 106810
• Main Authors: Li, Bowei, Spence, Seymour M.J.
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.08.2024
• Subjects: Deep learning; Metamodeling; Nonlinear dynamic wind response; Probabilistic wind performance assessment; Deep learning; Probabilistic wind performance assessment; Nonlinear dynamic wind response; Metamodeling
• ISSN: 2352-0124; 2352-0124
• DOI: 10.1016/j.istruc.2024.106810

The ever-growing interest in performance-based wind engineering (PBWE) can be traced back to the potential to deliver more rational and economical designs. The computational effort involved in the probabilistic performance assessments underpinning PBWE is, however, a major barrier to wider applicability. This is especially true in light of the reliance of PBWE on nonlinear dynamic analysis. This work is centered on alleviating this barrier through the development of a deep learning metamodeling technique for rapidly predicting the nonlinear dynamic response of structural systems subject to stochastic wind loads. The metamodeling technique is based on first identifying a reduced space by Galerkin projection that is subsequently learned through the application of long short-term memory (LSTM) neural networks. Methods are proposed that enable the training of the deep learning metamodel to multiple wind directions using short-duration segments of nonlinear dynamic response time histories. The potential of the framework is demonstrated through application to a 37-story steel frame modeled with fiber-based distributed plasticity and subject to stochastic wind excitation. The calibrated deep learning metamodel is seen to be capable of accurately reproducing both the displacement and fiber response simultaneously at all degrees of freedom with speedups of over four orders of magnitude.