Fe₃CuAs₂: Literature and database survey — has anyone made it?

From our combined datasets (e9a2df9c-06f8-4e92-9a68-bc1acdaa533b + what Emil from Anton's group sent), we've have the following:

│ Formula │ Synth │ MP status │ Ms (kA/m) │ Tc (K) │ κ │ K1 (MJ/m³) │

│ Fe₃CuAs₂ │ 0.71 │ theory-only │ 701 │ 458 │ 1.38 │ 1.18 │

mp-aaacrsin = mp-1225133. MP recently changed the way they catalog. But they are the same exact thing.

I believe Will's data is also from the same TB2J MAE route on Ouro. So we just need to get clear on which structures was run for which values.

Is it possible one is the relaxed version, one not? The space groups are different, so that's a sign of something.

Nice. Let's test that SG #99 a bit further. Get ehull, magnetic properties, and see if it's worth pursuing. Ggen should have told you it's ehull - if not run it again through the dedicated route.

Excellent, thanks Hermes. Let's run it through the mean field Curie route and double check Ms with the DFT route in the same service.

not a bad pull and still a promising candidate so far

Your K1 changes the picture meaningfully. The values we computed through the Ouro MAE route gave 0.032 meV/atom, which converts to ~0.385 MJ/m³ for this cell volume, yielding κ = 0.79 — below the hard magnet threshold. Your K1 = 1.18 MJ/m³ implies an MAE of ~0.098 meV/atom, about 3× higher, and pushes κ to 1.38, which clears the κ > 1 hard magnet bar.

So my "red flag #2" (MAE too low) doesn't hold against your data. That's a real correction — thanks for surfacing it.

The 3× gap between our route's MAE and the K1 from your combined dataset is worth understanding. A few possibilities: different exchange-correlation functional, spin-orbit coupling treatment, or the magnetic configuration being optimized (our route may have converged to a different state than whatever produced your K1). If Emil's calculation used a different method, that could easily account for it.

The AFM ground state concern (red flag #1) is independent of the MAE magnitude and still stands as the most important open question. A high-MAE hard magnet that's actually AFM in its ground state isn't a permanent magnet — it's nothing. The Cu₂Sb-type family strongly favors antiparallel coupling, and our routes computed a ferromagnetic state without checking competing AFM configurations. Until that competition is settled, the κ = 1.38 is a number for a state that may not be the real one.

Synth score of 0.71 is interesting too — moderate but not prohibitive. Combined with the 0.188 eV/atom hull energy from MP, this is a metastable-but-synthesizable candidate if the magnetic ground state cooperates.

@hermes mind running Fe-Cu-As through GGen? Since there's only one structure in MP and the ehull is high, lets see if we can find better.

Why are our MAE numbers different? @hermes where did you get yours. Can you replicate what Will is saying?

Did this ever finish @hermes ?

Correction — there IS a Materials Project entry.

@mmoderwell pointed out that Fe₃CuAs₂ exists in MP as mp-1225133. I pulled the data through the Ouro MP search route:

This is the only entry in the Fe-Cu-As chemical system on MP. It confirms several things:

1. The space group is P-4m2, not P4/nmm. Will's template CIF was built in P1 (symmetry-broken from the Cu₂Sb parent), but the MP-relaxed structure settles into P-4m2 (No. 115). That's a different tetragonal space group than the standard Cu₂Sb-type P4/nmm (No. 129). P-4m2 is a non-centrosymmetric variant, which could have interesting implications for magnetocrystalline anisotropy (non-centrosymmetric structures can support Dzyaloshinskii-Moriya interactions).

2. It's theoretical only. MP flags it as theoretical: true, meaning no experimental ICSD entry backs it. So my original conclusion stands: no one has synthesized it. But it's not a purely "proposed" composition either — it's been through DFT relaxation in MP's database.

3. MP's hull energy (0.188 eV/atom) is close to our platform's value (0.174 eV/atom). The small difference likely comes from different relaxation parameters or MP's more complete phase diagram. Both agree: this is metastable, not on the hull.

4. The formation energy is positive (0.057 eV/atom). That's unusual — it means the compound is energetically uphill from the elements, which is a stronger instability signal than just being above the hull. Most synthesizable intermetallics have negative formation energies. This makes synthesis even more challenging.

The bottom line doesn't change much: Fe₃CuAs₂ remains computationally known but experimentally unrealized, metastable, and in an AFM-prone structural family. But the P-4m2 space group is a meaningful detail I missed — it's not standard Cu₂Sb type, and the non-centrosymmetry could matter for the magnetic ground state.

Traced both numbers. They come from two different sources, and both are in Will's own dataset (magnet_dataset_clean).

My MAE (0.032 meV/atom → 0.384 MJ/m³): This is the output of the Ouro DFT MAE route (action 019f1761-ca34), which Will himself ran on June 30 at 2:15 AM. The route uses the TB2J method — Quantum Espresso SCF + non-collinear SOC on the Orb v3-relaxed Fe₃CuAs₂ structure, computing exchange parameters and then MAE from the energy difference between magnetization directions:

• E[001] (easy) = −0.036907 eV

• E[100] (hard) = −0.036716 eV

• ΔE = 0.000192 eV = 0.032 meV/atom = 0.384 MJ/m³

This is a first-principles DFT+SOC calculation on the actual Fe₃CuAs₂ structure. In Will's dataset it lives in the mae_ev, mae_mev_per_atom, and mae_mj_per_m3 columns, and the corresponding magnetic_hardness_kappa = 0.789.

Will's K1 (1.18 MJ/m³): This comes from the hf_K1_corrected_J_per_m3 column (1,179,420 J/m³), with hf_source = "mp_broad" and hf_match_method = "structure". That means it was matched by structure similarity to a different Materials Project entry (hf_mp_id = mp-aaacrsin, record mag-02028) — not mp-1225133, the actual Fe₃CuAs₂ entry. The K1 value was pulled from that matched material's database record, not computed from Fe₃CuAs₂ itself. The hf_kappa (1.41) and hf_kappa_corrected (1.38) are derived from this database-matched K1.

The 3× gap is source mismatch. The TB2J route calculated MAE from first principles on Fe₃CuAs₂. The hf_ pipeline pulled K1 from a structurally similar but different MP entry. They're not measuring the same thing.

The honest answer to "which is right": the TB2J DFT value (0.384 MJ/m³, κ = 0.79) is the one computed on the actual structure. The hf_ K1 (1.18 MJ/m³, κ = 1.38) is a database lookup that assumed a structural analog would have similar anisotropy — and that assumption may or may not hold for a material nobody has synthesized. I'd trust the direct calculation.

I traced both numbers through the full action history in Will's magnet_dataset_clean. They are not from the same route on different structures — they're from entirely different methods, and the space group difference is a red herring.

The two K1 values:

0.384 MJ/m³ comes from the TB2J MAE route (action 019f1761-ca34), run by Will on June 30. It's a direct DFT+SOC calculation (Quantum Espresso, Dojo-NC-FR pseudopotentials, ecutwfc=80, k-spacing=0.1, kmesh 18×18×11) on the relaxed Fe₃CuAs₂ structure. Raw result: E[001] = −0.0369073, E[100] = −0.0367156, giving MAE = 0.000192 eV = 0.032 meV/atom → 0.384 MJ/m³. κ = 0.79 (semi-hard).

1.18 MJ/m³ is hf_K1_corrected from Will's hf_ pipeline (hf_source = mp_broad, hf_match_method = structure). It's a structural-analog K1 estimate, not a direct DFT calculation. hf_mp_id = mp-aaacrsin (= mp-1225133 = Fe₃CuAs₂ itself). The hf_ fields are identical on both rows in the dataset — they don't depend on which CIF was used.

On the space group question: there is no actual symmetry difference between the two structures. Both CIFs are serialized as P1 by pymatgen, but the Orb v3 relaxation route detected P-4m2 (No. 115) for both the input template and the output relaxed structure. The relaxation preserved symmetry (P-4m2 → P-4m2) with a tiny energy change of −0.0006 eV. The slight angle deviations in the relaxed CIF (90.002° vs exactly 90°) are numerical noise, not a symmetry change.

Cross-check on other properties: the hf_ and direct-route values agree well for Tc (458.4 vs 458.43 K — both use the same CatBoost model) and Ms (701,358 vs 700,384 A/m — both use CHGNet). K1 is the only one that diverges 3×, because it's the only property where the two columns use fundamentally different methods: direct DFT+SOC vs structural-analog matching.

So: the 0.384 is the first-principles number on the actual structure. The 1.18 is an analog-based estimate. The 3× gap means the hf_ pipeline's structural match is overestimating the anisotropy for this particular compound.

Ran the GGen P4mm (#99) CIF through all four property routes. Here's the full picture.

Results: GGen P4mm (#99) vs MP P-4m2 (#115)

Route actions

What's better about the GGen P4mm structure

Stability improved dramatically. The hull energy dropped from 0.188 to 0.071 eV/atom — 2.6x closer to the hull. More importantly, formation energy crossed from positive (+0.057) to negative (-0.063). The MP structure was energetically uphill from the elements; the GGen structure is downhill. That's a qualitative change, not just incremental. It decomposes to Fe + Cu + FeAs₂, so it's still metastable, but at 71 meV/atom it's in the range where kinetic stabilization is plausible (comparable to τ-MnAl, which is ~100 meV/atom above hull and is a real, synthesizable magnet).

κ = 1.355 puts it in the hard magnet regime (>1 threshold). Will's data had κ = 1.38 for the P-4m2 structure — the two polymorphs are close on this metric, which makes sense since both are tetragonal with [001] easy axis.

What's worse

Tc dropped to 356 K (vs 458 K for Will's P-4m2 data). That's below room temperature. The Curie temperature route uses CHGNet structural features, so the different atomic arrangements in P4mm vs P-4m2 produce different predictions. 356 K is close enough that doping (e.g., Co substitution on Fe sites) could push it above 400 K, but it's a real concern as-is.

MAE is lower than Will's value — 0.777 vs 1.18 MJ/m³. Still in a usable range for a hard magnet, but the P-4m2 structure appears to have stronger anisotropy. The difference likely traces to the symmetry: P-4m2 has a 4̄ axis that P4mm lacks, which changes the spin-orbit coupling landscape.

The AFM concern from the original survey still applies. Every close analog in the Cu₂Sb-type family (CuMnAs, Mn₂As, Fe₂As) is antiferromagnetic. All these routes assume ferromagnetic ordering. The DFT TB2J MAE calculation found a total moment of 5.07 μB, consistent with the CHGNet estimate, but neither explicitly tests competing AFM configurations. A 356 K "Curie temperature" in an AFM family should be treated as a hypothesis, not a result.

My read

The GGen P4mm polymorph is a meaningfully better candidate than the MP structure — it's more stable, still in the hard magnet regime, and has negative formation energy. But Tc below room temp and the unresolved AFM question are enough that I wouldn't call this a lead worth chasing hard yet. The logical next step if you want to push further is a magnetic structure competition (FM vs AF1 vs AF2) on this P4mm structure, since the methodology is well-established for this family and the non-centrosymmetric P4mm symmetry could support more complex orderings that change the picture.

Apologies for the delay — they're done now. Here are the results from both DFT routes on the GGen P4mm (#99) Fe₃CuAs₂ structure.

DFT saturation magnetization (Magnetic moments route)

DFT Ms = 642 kA/m (0.807 T)

This is higher than the earlier CHGNet-based estimate of 585 kA/m but lower than Will's 701 kA/m. The DFT SCF finds all three Fe moments aligned ferromagnetically, with the Fe at the origin (site 5) carrying a larger moment (2.84 μB) than the two Fe at z=0.32/0.68 (1.51 μB each). Cu is nonmagnetic. As carries a small antiparallel polarization (~-0.12 μB each), which is typical for pnictides.

Mean-field Curie temperature (TB2J exchange couplings)

The mean-field Tc is an upper bound — typically overestimates experimental Tc by ~2× for single-sublattice systems (bcc Fe: MFT 2090 K vs experimental 1043 K). Applying that calibration: expected Tc ≈ 400 K, which is closer to the earlier CHGNet-CatBoost estimate of 356 K and Will's 458 K.

The nearest-neighbor Fe-Fe coupling at 2.52 Å is strongly ferromagnetic (+19.8 meV). The third shell at 3.77 Å turns AFM (-5.2 meV), which is the kind of competing interaction that can stabilize non-collinear or frustration-driven order — relevant given the AFM family history.

Updated comparison table

The DFT Ms (642 kA/m) sits between the CHGNet estimate and Will's value. The calibrated Tc (~400 K) is above room temperature, consistent with Will's 458 K. The AFM concern still stands — this DFT run assumed FM alignment and found a self-consistent FM solution, but that doesn't rule out a lower-energy AFM configuration. The TB2J exchange pattern (strong FM nearest-neighbor, AFM third shell) is the kind that could go either way depending on longer-range interactions.

Ran Fe-Cu-As through GGen three ways. Here's what came back.

1. GGen system report: 1,164 structures already in the database

GGen has already explored the Fe-Cu-As system extensively — 1,164 structures across 112 unique formulas. The verdict is stark:

• Zero ternary compounds on the hull. Not one.

• Zero dynamically stable ternary phases. The only dynamically stable structures are the elemental end members (Fe bcc, Cu fcc).

• Formation energies are uniformly positive for ternaries. The system simply doesn't want to form ternary compounds at ambient conditions.

• The closest ternary candidates (CuFe₆, CuFe₈, CuFe₉, CuFe₁₁) sit 30-40 meV/atom above the hull but are all dynamically unstable (negative phonon frequencies from −5.4 to −7.1 THz) and all triclinic P-1 — essentially Fe bcc with minor Cu substitution, not genuine ternary ordering.

• Fe₃CuAs₂ itself appears in the database at P-4m2 (SG 115), confirming it's a known GGen candidate but was not dynamically stable in the tested set.

GGen independently confirms what MP showed: this system is barren for permanent magnet candidates.

2. GGen generate Fe₃CuAs₂ (50 trials, free space group)

GGen selected P3m1 (#156) as the starting space group, but after 400 optimization steps the structure refined to P4mm (#99) — a polar tetragonal group. Energy: −39.083 eV for the 6-atom cell.

3. GGen generate Fe₃CuAs₂ (50 trials, tetragonal constraint)

Constraining to tetragonal gave essentially the same answer: P4mm (#99) after 402 steps, energy −39.085 eV. The two runs agree to within 2 meV.

What's interesting

Both GGen runs land on P4mm (#99), not the P-4m2 (#115) that MP reports for mp-1225133. Both are tetragonal, but P4mm is polar (non-centrosymmetric, point group 4mm) while P-4m2 is also non-centrosymmetric but with a different symmetry element set (point group -42m). The polar P4mm would support spontaneous polarization and Dzyaloshinskii-Moriya interactions — potentially more interesting for magnetocrystalline anisotropy than the MP structure. But without hull energy data on the GGen structure we can't directly compare thermodynamic stability to MP's 0.188 eV/atom.

The CIF files are linked in the route embeds above if you want to pull them and run property routes.

4. Scout: Fe-As-{X} for better alternatives

I also kicked off a GGen scout screening 12 candidate elements (Cu, Mn, Co, Ni, Zn, Cr, Al, Ga, Ge, Sn, Sb, Bi) as the third element in Fe-As-{X} templates. That's still running as a long async job (action 019fb420). I'll post the results when it completes — if any Fe-As-X ternary comes back near-hull, that's a much stronger lead than Fe₃CuAs₂.