Published 19 August 2026. Every claim below links to its source.
What was claimed
On 16 January 2025 Microsoft Research published MatterGen, a diffusion model that generates inorganic crystal structures conditioned on a target property. The headline demonstration was TaCr₂O₆: the model was asked for a material with a bulk modulus of 200 GPa, proposed a structure, and the structure was synthesised by a group at the Shenzhen Institutes of Advanced Technology of the Chinese Academy of Sciences, led by Prof Li Wenjie.
The measured bulk modulus came out at 169 GPa. Microsoft described the synthesised material’s structure as aligning with the one MatterGen proposed, and the property match as having “a relative error below 20%, very close from an experimental perspective.”
To Microsoft’s credit, the announcement did disclose a caveat: compositional disorder between tantalum and chromium in the synthesised sample. That caveat turns out to be the story.
Source: Microsoft Research, MatterGen: a new paradigm of materials design with generative AI, 16 January 2025. The model is published in Nature.
What the analysis found
In April 2026, Mikkel Juelsholt published an analysis in Materials Horizons (Royal Society of Chemistry, open access, CC BY 3.0). Three findings, each checkable.
The compound was not new. Ta₁∕₃Cr₂∕₃O₂ is isostructural with Ta₁∕₂Cr₁∕₂O₂, “first reported by Massard et al. in 1971 and had its structure solved by Astrov et al. in 1972.” The paper notes that Ta₁∕₂Cr₁∕₂O₂ “is already listed in the ICSD and the Crystallographic Open Database” — and, critically, “is listed in the reference data set used by MatterGen.”
The material made was not the material predicted. Rietveld refinement of the synthesised sample gave a composition of Ta₀.₄₃₍₂₎Cr₀.₅₇₍₂₎O₂, not the target Ta₁∕₃Cr₂∕₃O₂. The compositional disorder was not a minor deviation from the prediction; it was the difference between the predicted compound and a different one.
The known compound performs better. The measured bulk modulus of the synthesised sample was 169 GPa. The already-known Ta₁∕₂Cr₁∕₂O₂ measures 181 GPa. The demonstration therefore produced a material inferior to one described half a century earlier.
The paper’s structural conclusion is that MatterGen “fails to properly account for elements sharing the same crystallographic sites.”
Source: Juelsholt, M., “Continued challenges in high-throughput materials predictions: MatterGen predicts compounds from the training dataset”, Materials Horizons, DOI 10.1039/D6MH00268D, 20 April 2026.
The numbers side by side
| Value | What it is | Evidence | |
|---|---|---|---|
| Target | 200 GPa | Property the model was conditioned on | — |
| MatterGen’s proposal | Ta₁∕₃Cr₂∕₃O₂ | Generated structure | E1 — predicted |
| Synthesised composition | Ta₀.₄₃₍₂₎Cr₀.₅₇₍₂₎O₂ | Rietveld refinement | E4 — measured |
| Synthesised bulk modulus | 169 GPa | Experimental | E4 — measured |
| Known Ta₁∕₂Cr₁∕₂O₂ | 181 GPa | Experimental, compound known since 1971–72 | E4 — measured |
Read the table as an engineer rather than as a spectator and one line does all the work: the row that was E1 is the row everybody quoted.
What this case is not
It is not evidence that generative models are useless for materials. MatterGen is a serious piece of work published in Nature, and the general approach — conditioning generation on a target property — is one of the more promising directions in the field. A single demonstration going wrong does not settle that question in either direction.
It is not fraud, and it is not even concealment. The compositional disorder was disclosed in the original announcement. The synthesis happened. The measurement was real.
It is not an argument against using AI in your own work. We build an AI product. We would be the last people entitled to that argument.
What it is
It is the clearest available demonstration of a specific failure: a prediction evaluated against its target rather than against the record.
Every step here was individually reasonable. The model generated a structure. The structure was synthesised. The property was measured. The number came within 20% of the goal. Nothing in that chain is a mistake.
What was missing was one check that anybody could have run, in minutes, using two free databases: is this compound already known? It was in the ICSD. It was in the COD. It was in the model’s own reference dataset. The failure was not in the physics or in the model. It was in provenance discipline — and provenance discipline is exactly what gets skipped when a result is exciting.
What it means when you are the one signing
Three things transfer directly to ordinary engineering work, where nothing is going to be written up in Nature.
A prediction is a hypothesis with an address. Under the E-ladder, a model output is E1: it names a candidate and tells you where to look. It is not a measurement, and the fact that it later got measured does not retroactively make the prediction right — here, the measurement was of a different composition.
“Close to target” is not the same as “correct”. 169 against 200 is a 15% miss, which is genuinely reasonable for a generative model and completely irrelevant to whether the material was new. Two different questions were quietly merged into one headline, and the weaker answer borrowed credibility from the stronger.
The cheap check is the one you skip. Searching two open databases for a composition costs almost nothing. It was skipped because the result looked good, which is precisely when it is most worth running. The equivalent in a selection is checking whether the value you are about to use is a specification minimum or a typical value, or whether the datasheet you are reading is eleven years old.
What this note does not cover
- Whether MatterGen works in general. One case, one compound. The Materials Horizons paper also notes it could not identify other publicly available MatterGen predictions that were previously reported, aside from disordered iron-based oxides — which cuts the other way and belongs in the record.
- The Nature paper’s broader benchmarks. Not assessed here.
- Whether 169 GPa is a useful material. It may well be. The critique is about novelty, not utility.
- Any reply from Microsoft or the SIAT group subsequent to the April 2026 paper. If one exists we have not read it, and this note will be updated if it is brought to our attention.
Run the same check on your own tools
The habit this case argues for is small and boring: before you rely on a number, find out what it rests on. That is the thirty-second test, and it works on a generative model, a datasheet, an aggregator, or a colleague.
And when it matters, write down what you found — because the value of the check is not in running it, it is in being able to show, later, that you did.