We propose a data-driven latent space dynamics identification method (tLaSDI) that embeds the first and second principles of thermodynamics. The latent variables are learned through an autoencoder as a nonlinear dimension reduction model. The dynamics of the latent variables are constructed by a neural network-based model that preserves certain structures to respect the thermodynamic laws through the GENERIC formalism. An abstract error estimate of the approximation is established, which provides a new loss formulation involving the Jacobian computation of autoencoder. Both the autoencoder and the latent dynamics are trained to minimize the new loss. Numerical examples are presented to demonstrate the performance of tLaSDI, which exhibits robust generalization ability, even in extrapolation. In addition, an intriguing correlation is empirically observed between the entropy production rates in the latent space and the behaviors of the full-state solution.