Two Small AI Models Split the Work of Materials Research
Ask one large language model to design a crystal and you are asking it to do two jobs at once. The first is chemistry: what will hold together, what is likely to be stable, what has already been tried. The second is closer to clerical work: open the structure generator, hand it the right settings, read what comes back, call the next tool. Both run on the same words, and the usual fix has been a bigger model.
A major challenge of today's large language models is their enormous scale, run on hundreds of billions of parameters, the numbers a model learns during training, and still struggle with both domain reasoning and tool coordination in materials science. Tongyu Shi, Xue-Feng Yu and colleagues at the Materials Artificial Intelligence Center of the Shenzhen Institutes of Advanced Technology, part of the Chinese Academy of Sciences, went the other way. Their system, MatBrain, published Sept. 10 in Nature Machine Intelligence, gives the two jobs to two separate models, neither of them large by current standards.
MatBrain is an agent in the current sense of the word: not a chatbot that answers a question about a crystal, but a system that calls software tools, reads what they return and decides what to do next.
Mat-R1, the analytical model, carries 30 billion parameters and does the thinking about chemistry. Mat-T1 carries 14 billion and works as the executive, orchestrating the tool calls: what to run, in what order, on what. Both were trained on material that the group assembled and has now released, Mat-252K-SFT, a set of worked examples, and Mat-20K-RL, used for a later round of trial-and-error training. Between them they are meant to cover the span of a materials scientist's work, generating structures, predicting properties and planning a synthesis, and the paper reports versatility across all three.
The team's own diagnostic for whether that division of labor is real sits in an entropy analysis. Entropy here measures how spread out a model's choices are at each step, how undecided it is about what comes next. Tool planning and analytical reasoning, the paper reports, produce distinct output-distribution profiles, calling this "a diagnostic signal consistent with the functional specialization of the two modules." That wording is doing real work. The two kinds of output are measurably different. That the difference is what makes the two-model design pay off is a reading of the signal, not something the signal settles.
The demonstration is catalyst design, and specifically catalysts for bio-inspired nitrogen fixation: pulling nitrogen out of the air into a usable compound, the chemistry behind fertilizer. Set to that task, the authors report, MatBrain generated 30,000 candidate structures and identified 38 promising materials within 48 hours, "while significantly reducing the human-active time required for materials design and computational screening."
MatBrain, the authors write, "is competitive with frontier large language models while being lightweight and locally deployable." The journal received the paper on Feb. 12 and accepted it on July 30, after peer review. The authors declare no competing interests, and the funding is public money, from national and provincial research programs in China.
The code is archived on Zenodo under a CC BY 4.0 license, and both training sets are public on HuggingFace, which means a group with its own hardware can pull the pieces down and try the comparison.
Running locally is the whole point of building small: nothing goes out to an external service, nothing metered by the call. The main claim is therefore modest: matching the performance of larger models while remaining highly portable. For a materials group with a question, a rack of its own and no budget to rent somebody else's, what stands out is less the size of the search it ran than the size of the models that ran it.