AI Learns to Design Materials by Predicting Their Electrons, Not Just Their Atoms

For centuries, the discovery of new materials has been an exercise in patience. Metallurgists, chemists, and physicists mixed ingredients, fired furnaces, and tested samples one by one, hoping that somewhere in the combinatorial haystack a compound with the right blend of strength, conductivity, or optical behavior would emerge. In the last few years, artificial intelligence has promised to compress that timeline dramatically, and now a team at the Institute of Industrial Science, The University of Tokyo has taken a step that could make that promise far more concrete. The researchers have extended a leading generative AI model so that it designs not only the arrangement of atoms in a new material but also the behavior of its electrons at the same time. The new model, named MatterGen-e⁻, is reported in the peer-reviewed journal npj Computational Materials, and its central insight is deceptively simple: if you want a material with specific properties, you should teach the machine to think about the electrons first.

The principle driving the work is one of the oldest in materials science: structure dictates function. The hardness of a ceramic, the electrical conductivity of a metal, and the optical behavior of a semiconductor all trace back to two atomic-scale characteristics. The first is the crystal structure, the repeating three-dimensional pattern in which atoms or molecules arrange themselves. The second is the electronic structure, the way electrons are distributed among energy levels and around individual atoms within that lattice. Traditional approaches to computational materials discovery have focused almost exclusively on the first of these, generating candidate atomic arrangements and then screening them for desirable macroscopic properties. The Tokyo team argues that this is an inefficient way to bridge the gap between what a material is made of and what it can do, because the electronic structure sits physically in between, determined by the atomic arrangement and in turn determining the property of interest.

The new model builds directly on MatterGen, a generative model for inorganic crystals reported by Microsoft Research in early 2025. MatterGen is a diffusion-based model, a class of generative AI that learns to create new examples by gradually refining random noise into structured output, in the same general family as the image generators that have transformed visual computing. In MatterGen’s case, the output is not a picture but a stable crystal, described by three sets of variables: the types of atoms involved, their coordinates in space, and the geometry of the unit cell lattice that repeats to fill the crystal. Trained on existing crystal data, such a model can propose entirely new compounds that no chemist has synthesized. The catch, as the University of Tokyo researchers point out, is data. The amount of available information linking a macroscopic property directly to a crystal structure is limited, and it is difficult for a model to learn that correlation reliably from what exists.

Their solution is to give the model an intermediate language to work in. Rather than forcing it to leap directly from atomic coordinates to properties like band gap or formation energy, the team incorporated descriptors of the local electronic structure alongside the crystal structure itself. Two descriptors play the central roles. The first is the Bader charge, a well-established quantity from computational chemistry that assigns an electronic charge to each atom in the crystal lattice, effectively capturing how electrons are shared or transferred between atoms. The second is the atomic density of states, which describes how the available electronic energy levels are distributed around each atom. By generating these electronic descriptors jointly with the atomic structure, the model gains a physically grounded stepping stone between the arrangement of atoms and the properties that arrangement produces.

Lead author Ibuki Okuda frames the logic clearly. The electronic structure, he explains, physically determines a macroscopic property and is strongly constrained by the local atomic arrangement. Consequently, the electronic descriptors serve as an intermediate representation between structure and property. This is more than a convenience; it is a hypothesis about how machine learning should be organized for materials design. Instead of asking a neural network to find a statistical shortcut across a data-scarce landscape, the researchers are embedding known physics into the architecture of the generative process itself. The model does not have to rediscover from scratch that electron distribution matters. It is built to reason through it, which means every training example carries more usable signal than it would in a purely structure-based model.

The practical results support the hypothesis. When the researchers specified target values for formation energy and band gap, two fundamental quantities that determine whether a crystal is stable and how it behaves as an electronic material, MatterGen-e⁻ produced substantially more structures that were simultaneously valid, stable, unique, and novel than the original MatterGen managed. Those four criteria matter enormously in real materials discovery. A candidate structure that is valid but unstable will never exist outside a computer; one that is stable but a duplicate of a known compound offers nothing new. Generating candidates that clear all four hurdles at once is the benchmark of a genuinely useful generative model, and the electronic descriptors measurably raised the bar. As a bonus, because the model outputs electronic descriptors together with each structure, it also provides the density of states of newly generated materials, and those predictions agreed reasonably well with first-principles calculations, the computationally expensive quantum mechanical simulations that traditionally serve as the gold standard.

Perhaps the most convincing evidence comes from the control experiments. Senior author Teruyasu Mizoguchi describes how the team replaced the physically meaningful electronic descriptors with noisy Bader charges, or with a dummy variable carrying no physical meaning at all. The result was telling: the stability and validity of the generated structures improved only when physically meaningful descriptors were used. The quality of the electronic information, not the number of variables, is what matters. This finding cuts against a common temptation in machine learning, where adding parameters and inputs is often assumed to help. Here, the benefit flows entirely from the physics embedded in the representation. A meaningless variable of the same dimension contributes nothing, while a physically grounded one transforms the model’s ability to navigate toward stable, novel crystals. For a field increasingly awash in data-hungry black-box models, that is a design principle worth broadcasting.

The implications extend well beyond bulk crystals. Because the electronic descriptors are defined for each individual atom, the approach is not limited to periodic three-dimensional lattices. Local electronic states govern the properties of surfaces and interfaces, the regions of a material where bulk behavior breaks down and where phenomena like catalysis, corrosion, and charge transport are decided. These are precisely the environments that are hardest to model and most valuable to engineer, from battery electrodes to semiconductor junctions to catalyst surfaces. The Tokyo team suggests that using this model could improve the generation of surfaces and interfaces, opening a design space that structure-only generative models struggle to explore meaningfully. If the electronic-structure-first philosophy holds in those settings, it could reshape how researchers approach some of the most technologically consequential problems in energy and electronics.

The work also fits into a broader shift in how science thinks about inverse design. Instead of the slow trial-and-error of laboratory synthesis or the computationally expensive cycle of randomly generating candidates on a computer and screening the results, inverse design specifies the target property first, such as a particular band gap or formation energy, and asks a generative model to produce a matching stable crystal directly. Generative machine learning models are well suited to this task because they learn the statistical structure of existing crystal data and can then create entirely new materials that satisfy the specification. What MatterGen-e⁻ adds to that picture is a physically informed pathway: property to electronic structure to atomic arrangement, mirroring the causal chain that nature itself follows. The result is a generative model that does not just imitate the database of known materials but reasons about why those materials behave the way they do, and uses that reasoning to invent new ones.

None of this means the era of the laboratory is over. Generated candidates must still be synthesized, characterized, and validated in the real world, and first-principles calculations remain essential for confirming predictions at the highest accuracy. But the direction of travel is unmistakable. By building the physics of electrons into the heart of a generative model, the University of Tokyo researchers have shown that the fastest route to a material with the property you want may run through the energy levels you cannot see. As generative AI continues its march through chemistry and materials science, the lesson of MatterGen-e⁻ is likely to echo: the models that succeed will not be the ones with the most parameters, but the ones that understand, in some genuine sense, what they are building.

Subject of Research: Generative AI model that jointly designs crystal structures and electronic structure descriptors for inverse materials design

Article Title: Using electronic structure to guide AI in designing new materials

Article References: Using electronic structure to guide AI in designing new materials. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: generative AI, materials design, electronic structure, crystal structure, inverse design, MatterGen, diffusion models, band gap, Bader charge, density of states, npj Computational Materials, University of Tokyo