Meta’s Open Source AI for Science Summit brought together researchers working across chemistry, neuroscience, materials science and mathematics. The announcements show how AI research is expanding beyond chatbots into tools that simulate physical systems and help scientists work with large scientific datasets.
Three developments stood out: UMA Nitro, an expanded target for the Digital Brain Project and new ways to bring chemistry tools into classrooms.
UMA Nitro: a faster model for atoms
Meta announced UMA Nitro, a faster version of its Universal Model for Atoms. Models in this family are designed to help simulate atomic systems. Such simulations can help researchers study how materials behave and explore candidate structures before committing time and resources to physical experiments.
Meta described work with Carnegie Mellon University and the University of California, Riverside, in which UMA models identified crystal forms of drug molecules without additional training data. The crystal form of a drug can matter because it influences properties such as how a medicine dissolves.
Meta also described work with the University of Chicago involving simulations of enzymes at the atomic level. Enzymes are proteins that help chemical reactions happen in living systems. Simulating their structure and behaviour may help researchers understand how they work and explore possible applications in medicine, energy and sustainability.
These are research directions and reported study results. They should not be read as proof that an AI model can independently design a finished medicine or industrial material.
Digital Brain Project expands its data target
Meta said the Digital Brain Project is increasing its target from 15,000 hours of de-identified brain data to 25,000 hours. The project was launched with the Rothschild Hospital Foundation and the University of Montreal.
The aim is to support researchers studying the brain and human intelligence. More data can help research teams examine patterns across a wider range of observations, but the quality, diversity, consent arrangements and governance of the data matter as much as the total number of hours.
De-identification is one privacy safeguard. It does not remove the need for strict access controls, ethical oversight and clear rules about what researchers can do with the data.
Bringing chemistry into classrooms
Meta also highlighted the use of its scientific models in education. The summit included work to make UMA Playground available to high school chemistry teachers. The broader idea is to let students explore scientific concepts through interactive tools rather than relying only on static diagrams and textbook explanations.
For schools, the opportunity is practical only if the tools are accessible, aligned with the curriculum and supported by teachers. A simulation should help students understand a concept, not become a black box that produces answers they cannot explain.
Why this matters for India
India has a large base of pharmaceutical researchers, engineering colleges, science universities and school students. Open scientific models and accessible tools could lower the cost of experimentation for institutions that cannot build every simulation platform from scratch.
But access to a model is only one part of the problem. Institutions need computing resources, trained researchers, reliable datasets, domain expertise and ways to validate outputs against experiments.
For Indian universities, a useful starting point would be a small, supervised project: choose one chemistry or materials science question, compare model predictions with published or laboratory results, and document where the model fails.
What to watch next
The next meaningful signals will be independent reproduction of the research results, published evaluation details, access to the tools and evidence that researchers can use them reliably. A faster model is valuable when it enables more useful experiments, not simply when it completes a benchmark faster.
Meta’s summit points to a broader shift: AI is becoming part of the scientific toolkit. The hard work remains the same. Good science still needs evidence, reproducibility and human scrutiny.
Source and reporting note: This article reports announcements and research examples shared by Meta at its Open Source AI for Science Summit. We were not involved in the research or experiments. Please consult Meta’s original post and the linked institutional research for methods, study details and limitations.