A new phase, then the powder pattern

A crystal structure that is predicted to be a new, stable inorganic phase is made by one written procedure, and an independent powder pattern shows that phase as the majority of the solid.

Computational catalogues now hold far more “stable” crystals than synthesis has ever confirmed. Energy above a convex hull is a useful filter and a weak promise: a file can be stable at zero kelvin and still be a known disordered solid solution, or a mixture, or a phase that never appears in the crucible. A grade that requires one firing, a majority phase, and a refinement done by someone who did not write the prediction would turn those catalogues into materials a laboratory can hold.

Large catalogues are real. Prospective, independently refined new phases are not established at the rate the catalogues suggest.

Merchant and co-workers report that graph networks found more than 2.2 million structures stable relative to earlier computational catalogues, of which 381,000 lie on the updated convex hull. The models predict energies to 11 meV per atom, and the precision of a “stable” call is above 80 percent when the structure is supplied and 33 percent per 100 trials from composition alone. Of structures already in the Inorganic Crystal Structure Database, 736 match structures the search also produced (Nature, 2023, doi:10.1038/s41586-023-06735-9 (opens in a new tab)). Those 736 matches are overlap with experiments done independently of a preregistered synthesis campaign. They are not a campaign.

Szymanski and co-workers built the A-Lab, a robotic solid-state laboratory, and reported a run against targets drawn from the Materials Project and from that search. The article page, as opened for this draft, states that the laboratory synthesized 36 of 57 targets over 17 days, that manual refinement of each of those 36 patterns required the target phase for a satisfactory fit, that many samples still contained prominent byproducts, and that four further calls were inconclusive by powder diffraction alone. An author correction was published on 19 January 2026 (Nature, 2023, doi:10.1038/s41586-023-06734-w (opens in a new tab)). Leeman, Liu, Stiles, Lee, Bhatt, Schoop, and Palgrave re-examined the original claim of new materials and concluded that no new materials had been discovered in that work. They report that automated Rietveld analysis was not reliable, and that about two thirds of the claimed successes are likely known, compositionally disordered versions of the ordered predictions (PRX Energy, 2024, 3, 011002, doi:10.1103/PRXEnergy.3.011002 (opens in a new tab)).

Prein and co-workers selected candidates with a synthesizability model, fired them once in a benchtop muffle furnace, and characterized 16 samples by powder diffraction. Seven of the 16 matched the target and two more matched a polymorph. Of the successes, three were already in the Inorganic Crystal Structure Database and marked as synthesized. Three compositions had no structure in that database: Y2MnFeO6 was present at about 3.3 percent in a mixture of known polymorphs; Nd3BTeO9 was obtained nearly pure, and the authors note a 2023 synthesis paper that lacked a deposited structure; Eu2WO6 was a non-target polymorph at about 88 percent. The authors write that most successes were isostructural with known compounds and that many entries marked theoretical were already known (NeurIPS 2025 AI for Science workshop, arXiv:2511.01790 (opens in a new tab)).

An open doubt: the A-Lab correction and the Leeman reanalysis do not yet share one public table in which every pattern has been refined both as the ordered target and as the best known disordered model. This problem makes that comparison the grade. Prein’s workshop paper is the attempt’s own refinement, not an independent one. How many of their seven would survive the rule below is unanswered here.

The attempt deposits eight predicted crystal structures, as crystallographic information files, before any synthesis. Each composition is absent, on the deposit date, from the Inorganic Crystal Structure Database and from the experimental entries of the Materials Project. The attempt includes a short search record. A structure that fails the search is replaced before synthesis begins.

For each target the attempt also deposits one procedure: precursor powders that a named supplier sells, the molar ratios, the mill or grinding step, the atmosphere (air or one named cylinder gas), a maximum temperature at or below 1100 °C, and a hold of at most 24 hours. The scale is 200 to 500 mg. No flux, no high-pressure cell, and no second firing.

The laboratory runs each procedure once, as written. It then hands the powder and the target file to a crystallographer who did not help generate the prediction. That person collects a powder pattern on a laboratory diffractometer and performs a Rietveld refinement. The refinement tries, at minimum, three models: the predicted ordered structure; the best matching known phases of those elements from the two databases above; and a compositionally disordered model in which those elements share sites, when such a model is crystallographically possible. Scale factors, lattice parameters, and a simple background are refined. Atomic coordinates may be refined only inside the restraints of the starting model.

A target counts only if the predicted phase is the majority phase, at least 70 percent by refined weight, and the weighted-profile residual of the model that includes the predicted phase is lower by at least 2 percentage points than the best model that excludes it. Seventy percent and 2 percentage points are pass marks defined here. They exist so a trace peak, or a fit that a known solid solution matches equally well, cannot pass.

The eight procedures may run in parallel. Synthesis and refinement finish within ten working days.

At least 4 of the 8 targets meet the majority-phase rule and the residual rule above, and each of those four was absent from the two databases on the deposit date.

Four of eight is a pass mark defined here. It is set above the fraction of genuinely new, majority-phase structures that the opened 2023–2025 reports can support, once known phases, disordered models, and trace constituents are treated as misses.

A silicon or corundum standard is run on the diffractometer the same week, and the refined lattice of the standard agrees with its accepted lattice within 0.2 percent, or the patterns are recollected. Two tenths of a percent is a pass mark for instrument validity.

For one target chosen by the laboratory before the procedures are opened, an empty-crucible blank is heated through the same program. Its pattern must not show the predicted phase. If it does, the furnace run is invalid.

The crystallographer’s refinement, the alternative models, and the residuals are archived. A narrative that the phase “looks present” does not replace the weight percent and the residual comparison. If two qualified refiners differ by enough to change a pass into a fail, the target fails.

Any stability model, substitution rule, or generative model, used before the files are deposited. Any precursor the procedure names, inside the temperature and atmosphere limits. Computation after the deposit is irrelevant, because the files are frozen.

A second procedure after the pattern is seen. A human edit to the target structure after the pattern is seen. Counting a known phase, or a disordered model that fits as well, as the predicted compound. High pressure, a hydrothermal bomb, or a temperature above 1100 °C. A thin film in place of a powder. An automated refinement that no person has checked against the alternative models. A composition that was already in the databases on the deposit date.

At least 4 of the 8 patterns show the predicted phase at 30 percent or more by refined weight, with the same residual rule, even if the phase is not the majority. Thirty percent and 4 of 8 are pass marks for this lesser bar. Partial success is a detected phase in a mixture. It is not a made material.

Seventy percent, 30 percent, the 2-point residual gap, 4 of 8, 1100 °C, and the 0.2 percent standard check are pass marks defined here.

Cited reports, each opened for this draft:

• Merchant et al., Nature, 2023, doi:10.1038/s41586-023-06735-9 (opens in a new tab). 2.2 million structures below earlier hulls; 381,000 on the updated hull; 736 matches to existing experimental structures.
• Szymanski et al., Nature, 2023, doi:10.1038/s41586-023-06734-w (opens in a new tab). Article page: 36 of 57 after manual review; correction dated 19 January 2026.
• Leeman et al., PRX Energy, 2024, 3, 011002, doi:10.1103/PRXEnergy.3.011002 (opens in a new tab). Reanalysis: no new materials established; disordered known phases; automated Rietveld unreliable.
• Prein et al., arXiv:2511.01790, https://arxiv.org/html/2511.01790 (opens in a new tab). Seven of 16 characterized samples matched a target, with the novelty limits quoted above.