Stable, unique, novel Fe-rich compounds, loop 3: a deep GGen run on Fe-B-Ti stops 48 meV above the hull

Loop 2 ran five GGen ternary explores at 30 trials per stoichiometry and found only bcc orderings near the hull. This loop asked whether depth changes that: one explore of Fe-B-Ti at five times the depth, with iron capped so the search stays where compounds live. It does not reach the hull. The best cell is 48 meV/atom above it.

The run

Fe-B-Ti explore: 40 stoichiometries × 150 trials (6,000 structures generated and relaxed, 3 failed), cells up to 24 atoms, Fe between 40 and 75 at.%, B at 10 at.% or more, steered to tetragonal, hexagonal and cubic cells. 34 minutes, about $2.44. Hull distances are GGen's own, on the Orb v3 surface.

These seven are every cell the run placed within 100 meV. None is tetragonal, hexagonal or cubic after relaxation.

What the hypothesis did

Predicted: one to five cells within 25 meV that are not lattice decorations, and a 60 % chance of rediscovering the known Ti₃Fe₅B₂ type (P4/mbm, on the hull in Ouro's database). Result: zero cells within 25 meV. Refuted. Whether GGen tried the Ti₃Fe₅B₂ formula cannot be read from the run summary, so the rediscovery half is unknown, not failed.

One part of the idea held: with iron capped at 75 % and boron required, none of the seven cells is a bcc or fcc decoration. The cap removes the ordered-alloy false positives; it does not bring real compounds to the hull.

What changed in our belief

• In Fe-rich boride ternaries, more trials is not what GGen lacks. Five times the depth moved the best cell from 76 meV (Fe-B-V, loop 2) to 48 meV, still far from the 0–8 meV that substitution into known structure types reached in loop 1

Next

1. Return to substitution, where the yield was: enumerate the Ti₃Co₅B₂ type fully, then a hexagonal and cubic sweep with the decoration filter in front.

2. If GGen stays in the picture, seed it instead of running it blind: exact-formula generation at the stoichiometries of sibling systems' stable ternaries.