Scientists from the Institute of Industrial Science at the University of Tokyo have modified a generative artificial intelligence (AI) model that creates new crystalline materials with specified properties. The model now takes into account not only the arrangement of atoms but also the electronic structure - the distribution of electrons across energy levels. The results were published in the journal npj Computational Materials.
Traditional search for new materials follows two paths: slow trial-and-error experiments or random generation of candidates on a computer followed by verification, which requires large computational resources. The inverse problem is promising: first a desired property is specified (for example, the band gap - the energy barrier that determines electrical conductivity, or the formation energy), and the AI selects a structure that has this property. Machine-learning generative models trained on existing crystal data are well suited to this task because they can create entirely new materials.
The properties of crystals - hardness, electrical conductivity, optical behavior - depend on the atomic scale: the crystal structure (the arrangement of molecules in a repeating three-dimensional pattern) and the electronic structure. In early 2025, Microsoft Research presented the generative model MatterGen for inorganic crystals. It is a diffusion model: it generates new stable structures using atom types, their coordinates and lattice parameters. However, with a limited amount of data, it is difficult for the model to capture the relationship between a macroscopic property and the crystal structure.
The University of Tokyo team made training easier by adding local electronic-structure descriptors - the Bader charge (the electronic charge assigned to each atom in the lattice) and the density of states (a characteristic of the distribution of electronic levels). "The electronic structure physically determines the macroscopic property and is strongly constrained by the local arrangement of atoms," explains lead author Ibuki Okuda. "Therefore our electronic descriptors serve as an intermediate representation between structure and property."
Given specified values of formation energy and band gap, MatterGen-e- created significantly more structures that were simultaneously valid, stable, unique and new than the original MatterGen. In addition, the model outputs the density of states for each new material, and these data agree well with first-principles calculations (quantum-mechanical calculations without fitting parameters).
"We conducted control experiments by replacing the electronic descriptors with noisy Bader charges or a dummy variable with no physical meaning," says senior author Teruyasu Mizoguchi. "The stability and validity of the structures improved only when physically meaningful descriptors were used. This means that the quality of the electronic information matters, not the number of variables."
The authors formulate a principle for generative models in materials science: use physically meaningful intermediate representations, such as electronic structure, and not only atomic structure. Because the descriptors are defined for each atom, the approach is not limited to bulk crystals - it can be applied to surfaces and interfaces, where properties are governed by local electronic states. For now these are research results, not a finished commercial product.