This loop asked where a GGen ternary search is most worth pointing, and tested one answer. We scored all 231 Fe–A–B systems (A, B from 22 cheap elements) by a "sibling gap": how many stable ternaries the chemically neighbouring systems hold, minus how many the Fe system holds. Then we ran five GGen explores to see whether the score predicts yield. It did not: no explore returned a new compound on a true structure type within 25 meV/atom of the hull.
Each explore: Fe at 50 at.% or more, all three elements required, up to 20 atoms, 60 stoichiometries × 30 trials, steered to tetragonal, hexagonal and cubic cells. Hull distances are GGen's own, on the Orb v3 surface.
System | Why chosen | Near-hull cells (≤ 100 meV) | ≤ 25 meV | Of those, true structure types | Best cell |
|---|---|---|---|---|---|
gap 2.0 | 3 | 0 | 0 | VFe₆B₂, Cm, 76 meV | |
gap 1.5 | 2 | 0 | 0 | ZrFe₁₅Sn, Cm, 88 meV | |
gap 2.75 | 33 | 12 | 0 | Ti₇AlFe₈, R-3m, on the hull | |
empty system | 27 | 4 | 0 | GaFe₁₇Si₂, C2/m, 9 meV | |
filled control | 6 | 0 | 0 | ZrFe₃Ge₂, P-3m1, 28 meV (a known phase) |
All 16 cells within 25 meV, including the two on the hull in Fe-Al-Ti, collapse to one-atom bcc iron when every atom is replaced by Fe: they are ordered decorations of bcc (B2 TiFe and Fe₃Si-like superstructures), not new structure types. In the two systems where real compounds are expected (borides, Zr stannides), nothing came within 75 meV.
Predicted: at least two of the three high-gap systems give a new true-type cell within 25 meV; the filled control and the empty system give none. Result: the control and the empty system behaved as predicted; the high-gap systems gave none. Refuted where it mattered.
GGen at this depth finds orderings on simple lattices easily and does not reach compound structure types in 9,000 trials across five systems. That matches loop 1
Five systems, one seed, 30 trials per stoichiometry. Our own earlier measurement is that 30 trials misses known ground states of 12–16-atom cells about a third of the time.
We did not check whether the stoichiometries of the known Fe-rich ternaries (for example Zr₃Fe₈Sn) were among the 60 sampled, so "GGen missed it" and "GGen never tried it" are not separated.
The score was built from an August copy of Materials Project and Alexandria without GNoME.
Test the score with the tool that worked: substitute into the sibling systems' stable ternaries to make the Fe analogues in the high-gap systems, and run them through the free hull funnel. If those land near the hull, the score is right and the generator is the limit.
One deep GGen run (150–200 trials, 32 atoms) on a single boride system, to see whether depth changes the answer before more breadth is bought.
"Near-hull and new" from an explore needs the same lattice-decoration filter as a substitution screen; without it Fe-Al-Ti would have looked like twelve discoveries.
The sibling-gap score may still be right about where compounds exist. This pilot tested the score and the generator together, and cannot separate them.
A score for where to search, tested with five GGen explores: no new compound on a true structure type within 25 meV/atom.