Advances in Generative Models for Accelerated Discovery of New Materials

ABSTRACT The discovery of new materials can drive tremendous social and technological progress. However, the vastness of the material space makes comprehensive exploration computationally infeasible. This paper reviews the inverse design methods of generative models in materials science, aiming to discover customized materials based on specific functional requirements. We focus on analyzing the structural characteristics, advantages, and challenges of three main types of generative models in material design: variational autoencoders (VAEs), generative adversarial networks (GANs), and denoising diffusion probabilistic model (DDPM). In addition, this paper discusses material design methods based on language modeling, using natural language processing (NLP) techniques to enhance the expression of material properties and the controllability of the generation process. Finally, it looks forward to the future development directions of generative models in materials science, emphasizing the importance of improving model efficiency, expanding generation capabilities, and enhancing the effectiveness of generated samples, which provides new ideas and methods for researchers in related fields.