A deep learning approach for reduced-order modelling of nonlinear structures

Modern engineering systems, such as complex space or civil structural systems, often exhibit high degrees of freedom, say, 105-106, which are typical examples of high-dimensional nonlinear dynamic systems. These systems pose a huge challenge for fast response prediction, which essentially requires low-dimensional or reduced-order modelling. This talk introduces a deep learning-based nonlinear model reduction algorithm, which leverages Autoencoder networks for feature extraction, representing dominant variables of nonlinear structure, and Transformer networks for modelling/prediction of evolution in the latent space, i.e., reduced dynamics. The trained model is computationally efficient compared to traditional brute-force simulation of multi-mode Galerkin models, and reliable results for different excitation frequencies also demonstrate its well generalization.