Microsoft Research

The directions that can currently be confirmed are: a generative model for inorganic materials published in Nature (MatterGen), an Earth-system foundation model published in Nature and now running in a shipping product (Aurora), a learned exchange-correlation functional that other people's chemistry packages have started to adopt (Skala), a protein-ensemble emulator published in Science (BioEmu), and an enterprise platform with ten named customers and no published result (Microsoft Discovery). Almost all of it can be downloaded and run. That is the distinguishing fact about this portfolio, and it is also why the sharpest counter-evidence on this page exists at all.

Scope note: this page tracks the public record only: the lab's own releases, the peer-reviewed literature, and third-party reporting. No inside information. It records how far public material goes, not how well the organization is doing.

Overview

1. Materials generation: MatterGen, and the paper that answered it

What can be confirmed: MatterGen was published in Nature on January 16, 2025. Structures it produces are "more than twice as likely to be new and stable" as those from previous generative models and "more than ten times closer to the local energy minimum." The model and its weights are public.

The paper's proof of concept is a single sentence, and it is the one that carries the weight: "As a proof of concept, we synthesize one of the generated structures and measure its property value to be within 20% of our target." The compound was TaCr₂O₆, made in disordered form as Ta₁/₃Cr₂/₃O₂.

That synthesis has been answered in print. Mikkel Juelsholt, Materials Horizons 13, 5672–5679 (2026), performs the crystallographic analysis and concludes that the synthesised phase "is not a novel compound but is identical to the already known compound Ta₁/₂Cr₁/₂O₂ reported in 1972 and actually included in MatterGen's training dataset." The stated mechanism is general rather than incidental: generative models struggle where several elements occupy the same crystallographic site, so a known disordered phase gets misclassified as a new ordered compound. The conclusion drawn is about workflow: the results "underscore the necessity of rigorous human verification in AI-assisted materials research, limiting their use for rapid and large-scale prediction of new materials."

This is worth stating precisely, because it is easy to overstate in either direction. The critique does not say MatterGen is worthless, and it does not touch the benchmark numbers. It says the one physical object in the paper was already known, and that it was in the training set. For a page that exists to separate "a model scores well" from "something new was made," that is the single most informative thing in Microsoft's materials record.

Boundary: generation and ranking of candidate crystals. Confidence: high that the model, the paper and the open code exist; low that the campaign produced a new material that was then made.

Open leads: whether the MatterGen authors reply in print; whether any independent group synthesises a MatterGen candidate and establishes it as new; whether the disorder failure mode gets handled at the model level rather than by human inspection.

2. Earth system: the one that reached production

What can be confirmed: Aurora was published in Nature on May 21, 2025: "a large-scale foundation model trained on more than one million hours of diverse geophysical data," which "outperforms operational forecasts in predicting air quality, ocean waves, tropical cyclone tracks and high-resolution weather, all at orders of magnitude lower computational cost."

Aurora 1.5 followed on July 9, 2026. It adds 22 variables to the original four, moves to hourly resolution, and adds probabilistic ensembles. It "outperforms ECMWF ENS on 88.9% of the evaluated variable-and-lead-time targets" and has "roughly one-third lower track error" for tropical cyclones than the original Aurora. Weights are on Hugging Face.

The detail worth holding onto is who built 1.5. It came from Microsoft Weather (the product team behind forecasts in Windows, Bing, Copilot, Edge and MSN, "reaching more than a billion devices across 180 countries"), extending a model that came out of the research lab. This is the only direction on this page where a research artifact has visibly crossed into routine operation, with a named enterprise user (BKW, for energy operations) attached.

Microsoft states its own limit, which is worth quoting because organizations rarely write this down: "Foundation models should complement—not replace—physics-based models and domain expertise."

Boundary: forecasting. Skill against an operational baseline is a real and checkable thing, and it is not a scientific discovery. Confidence: high (peer-reviewed publication, open weights, a named benchmark against a public operational system, and a shipping product).

3. Quantum chemistry: the part other people adopted

What can be confirmed: Skala is a neural exchange-correlation functional for density functional theory. Version 1.1, released August 20, 2026, reports a "weighted average error of 2.8 kcal/mol on GMTKN55" and ranks first in "32 of the 55 categories" of that benchmark, while on GPU having "the same cost as r2SCAN" (a semi-local functional). It was trained on "2.5x more data" than the first public release.

The decisive detail is not the benchmark, it is the adoption. Skala ships in CP2K and is being integrated into Psi4, FHI-aims, ORCA and VASP. Those are other people's codebases, maintained by other people, who have to support it. Among the companies tracked on this site, this is the rarest form of evidence: not a company reporting its own number, but independent maintainers deciding the thing is worth carrying. The stated limitation is small and specific: on CPU there is overhead for small systems, which "disappears for systems with more than ~300 orbitals."

Boundary: accuracy of a computational method. GMTKN55 is a test set, not a discovery; a functional that ranks first in 32 categories has not made anything. What it changes is the cost of everything downstream of it. Confidence: high (open weights, an external benchmark maintained by others, and third-party integration).

Open leads: which of the four in-progress integrations reach a release; whether any published result attributes a finding to Skala rather than to the group using it.

4. Biomolecules: BioEmu

What can be confirmed: published in Science on August 14, 2025 (volume 389, issue 6761), with 28 authors. BioEmu "emulates protein equilibrium ensembles by generating thousands of statistically independent structures per hour on a single graphics processing unit," integrates "more than 200 milliseconds of molecular dynamics simulations, static structures, and experimental protein stabilities," and "predicts relative free energies with 1 kilocalorie per mole accuracy compared with millisecond-scale MD and experimental data." The code is open.

The paper names what it captures: "cryptic pocket formation, local unfolding, and domain rearrangements": the motions that matter for whether a pocket a drug could bind actually opens.

Boundary: emulation. BioEmu is a cheap substitute for an expensive simulation, benchmarked against that simulation and against experimental stabilities. It is not a substitute for an experiment, and the paper does not claim it is. Confidence: high (peer-reviewed, open code, accuracy stated against both simulation and experimental data).

5. Enterprise layer: Microsoft Discovery

What can be confirmed: introduced May 19, 2025 and generally available June 2, 2026. It is the commercial surface (agents, simulation and HPC on Azure) through which the models above are meant to reach industrial R&D.

The demonstration everyone quotes is a PFAS-free immersion coolant for datacenters: a prototype found "in about 200 hours," where it "otherwise would have taken months, if not years," and synthesised "in under four months." The validation sentence is narrower than the headline: "We have already tested some of the primary properties of this material and they align to the AI predictions." Microsoft's own hedge in the same post is "While this project is only an experiment." The other number in that release belongs to a different project: a solid-state electrolyte screen with PNNL that "screened over 32 million candidates" and produced a battery candidate using "70% less lithium."

Who is named at general availability: Yale Engineering (redox flow battery molecules), Georgia Tech (prebiotic plausibility of histidine), PNNL (organic redox flow battery materials, autonomous laboratory), Ginkgo Bioworks (agents running on Ginkgo Cloud Lab), Causaly, Cambridge Consultants, Wiley, BHP (copper leaching, "in a matter of months instead of years"), Syensqo (heat-transfer fluids) and GSK.

Boundary: ten named customers, zero published results. Every entry above is a description of work under way, sourced to the vendor's own announcement. That is evidence about adoption, which is real; it is not evidence about discovery. Confidence: high that the platform and the named engagements exist; low for any scientific outcome attributed to them.

What could not be found

• No published reply from the MatterGen authors to the Materials Horizons critique, and no independent synthesis establishing a MatterGen candidate as new.
• No peer-reviewed result from any named Microsoft Discovery customer.
• No fuller characterisation of the PFAS-free coolant beyond "some of the primary properties."
• No headcount or budget for the lab; the public page names its managing director (Bonnie Kruft) and its UK, Netherlands and Germany presence, and no more.
• No stated retirement or continuation plan for Project S, the earlier Science Engine framing, relative to Microsoft Discovery.

Method and scope

Scope: only work where AI is applied to scientific research. General model capability, Azure as infrastructure and consumer products are excluded except where a science model reaches them, as Aurora does.

Confidence is assigned claim by claim, not to a whole direction. The labels:

• High: a public artifact, execution or adoption by an outside institution, or review by a named expert;
• Medium: an official release or company-reported figure: inspectable in detail, unverified outside;
• Low: a single-source claim, a vendor announcement about work in progress, or a claim contested in print with no reply.

Two operating rules: revisions are dated by version and are not treated as new projects; evidence that a project exists is not evidence that the technology has succeeded.

One remark specific to this company. Microsoft's portfolio is the most open of any tracked here: weights, code and benchmarks are downloadable for four of the five directions. Openness is why an outside crystallographer could examine the MatterGen synthesis at all and publish a correction in a peer-reviewed journal. The correction is genuinely bad news for one claim, and the fact that it could be written at all is the strongest structural thing on this page.

Primary sources

Peer-reviewed

• A generative model for inorganic materials design, *Nature* (2025-01-16)
• Continued challenges in high-throughput materials predictions: MatterGen predicts compounds from the training dataset, *Materials Horizons* **13**, 5672–5679 (2026)
• A foundation model for the Earth system, *Nature* (2025-05-21)
• Scalable emulation of protein equilibrium ensembles with generative deep learning, *Science* **389**, 6761 (2025-08-14)

Official releases

• Aurora 1.5: Extending open foundation models for weather and Earth-system applications (2026-07-09)
• Broadening access to Skala creates a faster path to predictive DFT (2026-08-20)
• Transforming R&D with agentic AI: Introducing Microsoft Discovery (2025-05-19)
• Announcing Microsoft Discovery general availability (2026-06-02)
• Microsoft Research AI for Science