Inverse design of semiconductor materials with deep generative models

Abstract

In the realm of materials design, effectively navigating the expansive chemical design space to discover materials with specific desired properties remains a formidable challenge. Here, we introduce an inverse design framework to generate thermodynamically stable semiconductor materials, utilizing existing data on decomposition enthalpies, synthesizability information, and band gaps. This framework encompasses a compositions generation model (VGD-CG) that integrates conditional variational autoencoders (VAE), generative adversarial networks (GAN), and a state-of-the-art diffusion model (DM), alongside a template-based structure prediction (TSP) approach. In our quest for deeper insights into the strengths and limitations of VAE, GAN, and DM in inorganic materials design, we embarked on a comprehensive comparative study using general performance indicators. Subsequently, to verify the practicality of VGD-CG, we employed this framework to delve into the N–Ga, Si–Ge, and V–Bi–O compositional spaces. Through further theoretical calculations, we successfully identified several potential semiconductor materials. Our results underscore the effectiveness of the combination of VGD-CG and TSP in navigating the vast chemical design space, facilitating the discovery of novel semiconductor materials.