Can existing public quantum datasets support a move to nickel superalloys?

EXPERIMENTS.md · lines 977–1022

E22 — Superalloys are outside the support of every public DFT dataset

Date 2026-09-12 · Question Can the environment be extended to nickel superalloys using existing open data? · Provenance literature and database review

Result — measured across Materials Project, Alexandria, OMat24, sAlex, MC3D and NOMAD:

• Ni + Al + Cr + Co: zero structures in any of them
• Cantor alloy (CoCrFeMnNi): zero in any of them
• Zero entries anywhere contain all of Ni-Al-Cr-Co-Ti-Ta-W
• NOMAD holds 1,307 Ni-Al-Cr-Co entries and none are FCC
• No disordered solid-solution sampling in any general corpus; they are ordered prototypes on small cells

A real superalloy is outside the support of every general-purpose DFT dataset. Extending there requires generating data, not harvesting it - the opposite of the refractory case, where RHEA already had exactly the right distribution.

Correction to the magnetism framing of E20. The concern was stated as a ~500 meV/atom ferromagnetic effect. For gamma/gamma-prime specifically that is wrong: its Curie temperature is 14 K, so it is paramagnetic across the whole 300-1200 K operating range. The real problem is not whether the alloy is magnetic at temperature but that reference DFT is computed with ferromagnetic initialisation, which may describe the wrong magnetic state entirely. CHGNet's role as a trust flag (E20) still holds, but the reason is different from the one given there.

Licence constraints relevant to an aerospace context.

• Meta FAIR checkpoints (eSEN, EquiformerV2-OMat24, UMA) carry an acceptable-use policy barring ITAR-subject and "transportation technologies" use. Worth legal review before any ESA-adjacent deployment.
• GNoME data: CC BY-NC 4.0. AFLOW: "free for scientific, academic and non-commercial purposes. Any other use is prohibited."
• FitSNAP training data is GPL-2.0 - software copyleft on data - and is incompatible with CC-BY-SA-4.0, so those sources can never be merged.
• Nitol CoCrFeNi and Wang Ni-Al datasets carry no licence at all.

Recommended path for superalloys: pretrain on OMat24 (CC BY 4.0), fine-tune on MatPES (BSD-3-Clause, spin-polarised PBE, no +U, metal-tuned smearing), then generate gamma/gamma-prime SQS and MC data directly using the Cao-Freitas and Sheriff recipes, both MIT-licensed with data.

Unverified. The exact FCC fraction of LeMat-Traj (112,932,152 rows) could not be computed - HuggingFace's filter index will not build over 419 GB. It is the only corpus carrying space group and n-ary as queryable columns alongside forces, so that number remains open.