<p>This paper leverages deep learning to find data-driven parametric low-dimensional latent space model/dynamics of infinite-dimensional nonlinear continuous structures, which favors an efficient prediction of frequency response curves (FRCs) with varying external excitations. Explicitly, an unsupervised learning algorithm is used to extract dominant variables in low-dimensional latent space via Autoencoder (AE), while a Transformer network is trained in the latent variable to capture their temporal evolution. Both forcing parameters and time domain (grid) variables are input of the training procedure, which makes the trained model distinct from previous studies: (1) enabling the latent model to perform well for unseen forcing scenarios, i.e., generalization across changes of forcing frequencies and amplitudes, (2) directly predicting long-time nonlinear responses (time series) without needing a recursive sequence prediction. Remind that a previous data-driven model for dynamics has to be recursively called to predict the next few steps using its current prediction as an input, meaning that a long-time sequence prediction (as required by frequency responses) is time consuming and also becomes less reliable over a long horizon. As two examples, low-dimensional parametric latent models for a nonlinear foundation beam and a sagged cable are found, upon which various frequency responses are efficiently built, demonstrating its superior performance over either the empirical Galerkin truncation or time-consuming brute-force simulation.</p>