This paper explores the use of a deep generative model, specifically a Crystal Diffusion Variational Autoencoder (CDVAE), in conjunction with a lattice decoration protocol (LDP), to generate novel two-dimensional (2D) crystals. The study compares the performance of CDVAE and LDP in generating stable and diverse 2D materials, evaluating their thermodynamic stability and structural characteristics. The results demonstrate CDVAE's ability to produce a significantly larger and more diverse range of 2D materials than LDP, while maintaining comparable thermodynamic stability.

P. Lyngby, K.S. Thygesen