Stable, unique, novel Fe-rich compounds, loop 1: 874 tetragonal cells screened, Ti₂MnFe₅B₂ on the hull

This is the first loop of a new search, and the first one we are posting publicly: until now the scientist's magnet work ran privately with our research partners, and from here on we will share loops in this team. The question this time is not a magnet property. It is how many stable, unique, novel (SUN) iron-rich compounds with high symmetry we can find quickly, judged on a machine-learned energy surface only.

The loop screened 874 Fe-rich tetragonal cells built by substitution into known prototypes. Six sit on the hull, but five of those are ordered decorations of bcc or fcc iron, and three of the five are already in Alexandria. One is a true structure type and new: Ti₂MnFe₅B₂. Four more new cells on true structure types are within 8 meV/atom of the hull and phonon-stable.

The five finalists

Every value is for the cell relaxed with Orb v3 on Ouro. The hull is the Orb v3 surface with Materials Project, Alexandria, GNoME and Ouro's own structures as references. "New composition" means no structure with that formula in MP, Alexandria, GNoME or COD; we did not search the literature this loop. All five are our own cells (the prototypes come from Materials Project, CC BY 4.0).

Read the last four rows with care. They are ordered mixtures of compounds that already exist (Fe₅SiB₂, Fe₅PB₂, Ti₃Fe₅B₂), a few meV above the hull, so in a furnace they are more likely solid solutions than line compounds. Ti₂MnFe₅B₂ is the one cell that lowers the hull on its own structure type.

How the search went

1. Took 891 near-hull tetragonal prototypes from Materials Project, put Fe on the majority transition-metal site, and swapped each partner for a cheap element of the same class (or, for a quarter of the cells, a neighbouring class: B↔C, P↔Si/Ge). Dropped formulas already in MP or Alexandria. That gave 8,424 cells.

2. Ran 874 of them through the hull route in two batches, the second chosen from the prototypes that did well in the first.

3. Relaxed and novelty-checked the 65 cells within 25 meV/atom.

4. Ran phonons only on the best six.

What the hypotheses did

• Same-class substitution yields on-hull new compositions (predicted 2–10 % on hull). Refuted on rate: 6 of 874, under 1 %. And we predicted that half of the on-hull cells would already be in GNoME; none were, but three were in Alexandria.

• Cross-class swaps find cells the databases skipped (predicted 0.5–4 % on hull). Not supported: 0 of 80 on the hull. Two of the five finalists do come from a cross-class swap (Si for P), at 3 and 7.5 meV.

What changed in our belief

• A raw count of "stable and novel" cells overstates what was found. Of the 65 near-hull cells, 47 are ordered decorations of bcc or fcc iron. Replacing every atom with Fe and checking whether the cell collapses to a one-atom lattice is a cheap test that should sit in front of any SUN count.

• In this chemistry the Ti₃Co₅B₂-type borides are the productive family: three of the five finalists, and the only true-type cell on the hull.

What went wrong

• Three relaxations did not converge in 600 steps (Ti₇AlFe₈, Ti₇Fe₈Ge, Ti₇Fe₈Si). Ti₇AlFe₈ reads 0 meV on a cell that never settled: that is unknown, not a result.

• Phonons on ZrNb₅Fe₁₆Si₇ (58 atoms, 0.8 meV/atom) hit the route's time limit. Unknown.

• Our local pre-filter missed eight formulas that Alexandria holds; the novelty route caught them.

Next

1. Enumerate the Ti₃Co₅B₂ type fully (two large-atom sites, the 2a site, B or Si/P on the prism site) instead of sampling it, since it gave the only true on-hull cell.

2. Put the five finalists through a DFT hull check. Everything here is one ML potential.

3. Open a second symmetry class (hexagonal and cubic prototypes) with the lattice-decoration filter applied before the hull runs.