Using electronic structure to guide AI in designing new materials

edited by Sadie Harley, reviewed by Robert Egan

This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility:

Add to Preferred Sources Researchers develop a generative AI model that designs crystal and electronic structures together. Credit: Institute of Industrial Science, The University of Tokyo Structure dictates function. Scientists are increasingly using artificial intelligence (AI) to design the arrangement of atoms in new materials with desired properties. Now, researchers at the Institute of Industrial Science, The University of Tokyo, have extended a leading generative AI model so that it designs the electronic structure of a material together with its crystal structure. The new model, named MatterGen-e⁻, is reported in npj Computational Materials.

New materials have traditionally been invented using trial-and-error experiments (a slow process) or randomly generating candidates on a computer and screening the results (which is computationally expensive).

A game changer is inverse design—a target property is specified initially (e.g., the specific band gap or formation energy), and AI rapidly and directly designs a matching structure for a stable crystal. Generative machine learning models, which learn from existing crystal data, are well suited to this task because they can create entirely new materials.

The chemical and physical properties of a crystalline material (e.g., the hardness, electrical conductivity and optical behavior) depend on atomic-scale properties: the crystal structure (the arrangement of molecules in a repeating three-dimensional pattern) and the electronic structure (how electrons are distributed among energy levels and around atoms).

Microsoft Research reported a generative model for inorganic crystals (MatterGen) in early 2025. This diffusion-based model generates new, stable crystal structures using structural variables: atom types, coordinates and lattices. However, it is difficult for a model to learn a correlation between a macroscopic property and the crystal structure using the limited amount of data available.

The team at the University of Tokyo made it easier for the model to learn from data by incorporating descriptors of the local electronic structure alongside the crystal structure—the Bader charge (the electronic charge assigned to each atom in the crystal lattice) and the atomic density of states (descriptors of the local electronic structure).

"The electronic structure physically determines a macroscopic property and is strongly constrained by the local atomic arrangement," says lead author of the study Ibuki Okuda. "Consequently, our electronic descriptors serve as an intermediate representation between structure and property."

For target values of the formation energy and band gap, MatterGen-e⁻ produced substantially more structures that were simultaneously valid, stable, unique and novel than MatterGen. By outputting electronic descriptors together with each structure, MatterGen-e⁻ also provides the density of states of newly generated materials, which agreed reasonably well with first-principles calculations.

"We performed control experiments replacing the electronic descriptors with noisy Bader charges or a dummy variable with no physical meaning," explains senior author Teruyasu Mizoguchi. "The stability and validity of the generated structures improved only when physically meaningful descriptors were used. Thus, the quality of the electronic information, not the number of variables, is what matters."

These findings point to a design principle for generative models in materials science: using physically meaningful intermediate representations, such as electronic structure, rather than atomic structure alone.

Defining electronic descriptors for each atom means the approach is not limited to bulk crystals. Using this model could improve the generation of surfaces and interfaces, where local electronic states govern material properties.

Inverse Materials Design via Joint Generation of Crystal Structures and Local Electronic Descriptors, npj Computational Materials (2026). DOI: 10.1038/s41524-026-02322-4 Journal information: npj Computational Materials

BSc Life Sciences & Ecology. Microbiology lab background with pharmaceutical news experience in oil, gas, and renewable industries. Full profile →

Bachelor's in mathematical biology, Master's in creative writing. Well-traveled with unique perspectives on science and language. Full profile →