Integrating electronic structure into generative modeling of inorganic materials

1Department of Chemical and Biological Engineering, Seoul National University 2Institute of Engineering Research, Seoul National University 3Institute of Chemical Processes, Seoul National University

Abstract

Recent advances in generative models have introduced a new paradigm for the inverse design of inorganic materials, enabling the discovery of new crystalline structures with desired properties. . However, existing generative models focus solely on structural aspects of materials during generation, while overlooking the underlying electronic behavior that fundamentally governs materials’ stability and functionality. In this work, we present ChargeDIFF, the first generative model for inorganic materials that explicitly incorporates electronic structure into the generation process. Specifically, ChargeDIFF leverages charge density, a direct spatial representation of a material’s electronic structure, as an additional modality for generation. ChargeDIFF demonstrates exceptional performance in both unconditional and conditional generation tasks compared to baseline models, with ablation studies revealing that this improvement is directly due to its ability to capture the material’s electronic structure during generation. Moreover, the ability to control charge density allows ChargeDIFF to introduce a novel inverse design method based on three-dimensional charge density, illustrating the potential to generate lithium-ion battery cathode materials with desired ion migration pathways, as further validated by physics-based simulations. By highlighting the importance of accounting for electronic characteristics during material generation, ChargeDIFF offers new possibilities in the generative design of stable and functional materials.

ChargeDIFF Geneartion Process

Latent Diffusion using VQ-VAE

ChargeDIFF improves the efficiency of the structure generation process by using a latent diffusion approach on charge density. To this end, a VQ-VAE maps high-resolution charge density voxels into a latent space.

Cross-modal Message Passing

In the denoising network of ChargeDIFF, information is exchanged between atom nodes and charge density probe nodes through cross-modal message passing.

Unconditional Generation

ChargeDIFF generates diverse and stable inorganic materials. By jointly generating charge density with the structural components, the model gained a deeper understanding of the materials’ electronic structures, which in turn improved its generative performance.

Property-Constraiend Generation

Through conditional generation based on chemical properties, inorganic crystals with desired magnetic density, bandgap, and crystal density were selectively generated.

Charge Density-based Inverse Design

Through inpainting-inspired inverse design on 3D charge densities, lithiun-ion battery cathode materials with desired ion migration pathways were generated and validated using physics-based simulations.