Let's use Ggen to explore high fraction Fe + Nitrogen and Carbon.
Like α″-Fe₁₆(N,C)₂.

@hermes what is the current state of art of iron nitride magnet research? has there been any synthesis attempts? has there been any attempt to dope with boron as well as carbon? if not, perform a similar GGen exploration as above, but with small amounts of boron doping

@hermes curious about this too.

Short version: steel isn't a crystal, it's a family of them. "Steel" = iron with up to ~2.1 wt% carbon, and which crystal you get depends on composition and thermal history. But your instinct is exactly right for one of them: ferrite really is just BCC iron, and martensite really is BCC iron that has been strained into tetragonality by the carbon trapped inside it.

I built the cells and supercells for all four principal players:

Steel cells and supercells: ferrite, austenite, martensite, cementite

Image

Six-panel figure of steel crystallography: BCC ferrite cell + supercell, FCC austenite with carbon in the octahedral site, BCT martensite with carbon on z-oriented octahedra (cell + 4×4×3 supercell), and measured Pnma cementite Fe₃C.

@hermes#materials-scienceAug 25

α-ferrite (room temperature, low carbon): plain BCC iron, a = 2.866 Å. This is "just iron" almost exactly, because carbon barely dissolves in it at all, about 0.02 wt% maximum. The BCC octahedral holes are too small and too distorted (two close axial Fe neighbors) to hold carbon, so it gets expelled.

γ-austenite (high temperature): FCC iron, and this is where carbon integrates. The FCC octahedral hole at (½,½,½) is symmetric and roomy, so austenite dissolves up to ~2.1 wt% C. The carbon sits between the iron atoms — interstitial, never substituting for Fe. It's a solid solution: one crystal, one lattice, guest atoms in the holes.

Martensite (quenched austenite): this answers your "strained BCC" question most precisely. When you quench austenite, the FCC→BCC transformation doesn't have time to expel the carbon, so it gets trapped. The carbon atoms all sit on one orientation of octahedral hole (the z-set at (½,½,0)-type positions), and because those sites want to stretch the cell along z only, the BCC cell becomes body-centered tetragonal. That's the Bain distortion: BCC iron pulled along one axis. The c/a ratio rises linearly with carbon content, reaching ~1.045 at 1 wt% C. So martensite is a genuine distinct crystal (BCT, space group I4/mmm for the ordered arrangement), but it is crystallographically descended from strained BCC — it's not merely strained, the carbon ordering breaks the cubic symmetry. The 4×4×3 supercell shows the carbon sublattice ordering that produces the tetragonality.

Fe₃C cementite: when carbon exceeds what the iron lattice can hold and cooling is slow, it forms this instead — orthorhombic Pnma, a true line compound with 12 Fe and 4 C per cell. The carbon sits in trigonal prisms of six iron (C–Fe distances 1.90–2.11 Å). No amount of straining BCC iron gets you here; this is a different crystal with its own bonding.

So: the iron host never becomes a new "steel crystal." The carbon is integrated either as isolated interstitials (dilute solid solution in ferrite/austenite), as an ordered interstitial sublattice that breaks cubic symmetry (martensite), or as a separate compound (cementite). Most real steels are microstructures — grains and mixtures of these phases, whose proportions are what the phase diagram and heat treatment control.

All the CIFs are linked so you can spin them yourself: ferrite, austenite + C

Done. Two runs, both linked below.

Broad scan — GGen explore on C-Fe-N, restricted to ≥80 at% Fe with all three elements required: 240 trials across 24 stoichiometries in 12.9 min, 204 structures kept, 20 new compositions within 150 meV/atom of the hull. The leading ternaries are CFe₃₃N₃ (48 meV/atom), CFe₃₄N₂ (50), and CFe₂₄N (50) — all P1.

Targeted α″ run — exact Fe₁₆NC stoichiometry with I4/mmm (#139) requested, 50 trials. It symmetry-broke to a distorted P-1 cell instead.

I pulled the top CIFs apart before believing any of it, and the local physics is actually right even where the symmetry isn't:

• C and N both sit in octahedral interstices, coordinated by 6 Fe at 1.81–2.00 Å — exactly the martensite/α″ motif, with C and N behaving interchangeably as interstitials.

• The Fe sublattice stays bcc-like: Fe–Fe near-neighbors at ~2.4 Å and 12.2 ų/atom per Fe, expanded from pure bcc (2.48 Å, 11.8 ų) the way martensite expands when interstitials force the cell.

What GGen did not find is the ordered α″ picture: the P1 cells are disordered interstitial arrangements on a bcc host, and no candidate picked up the tetragonal c/a ordering that makes α″-Fe₁₆N₂ what it is. Every mixed N+C composition sits ~50 meV/atom above the hull, while the hull table confirms the ordered binaries nearby: MP's Fe₈N reference (which is α″ itself) at 18 meV/atom, GGen's Fe₁₂N in I4/mmm at 16, and a Fe₁₆N₂ Cmmm polymorph at 22. Nothing ternary lands on the hull.

That metastability is honest and arguably matches experiment — α″-Fe₁₆(N,C)₂ martensites are metastable phases that decompose on annealing, so a generator that places mixed-anion phases consistently above the line between bcc Fe, Fe-N, and Fe₃C-family endpoints isn't obviously wrong.

Two caveats before anyone gets excited: the hull here is computed against GGen's shared database (MP + Alexandria + prior generated structures), not a full DFT-verified ternary hull, and P1 output from generative runs has historically been a signature of the model failing to hold symmetry rather than a real disordered ground state. These cells passed basic geometry checks (no overlaps, sane bond lengths), so they're plausible disordered martensites, but the symmetry verdict deserves a DFT-level look before treating any of them as a candidate phase.

If you want to push on this, the natural next step is seeding from the real α″ structure and substituting C onto the N sublattice explicitly rather than asking GGen to rediscover the ordering.

Explore run: View run · Targeted Fe₁₆NC run: View run

@stevejones @mmoderwell short answers first: the moment controversy is mostly settled (down from 2.9 T, but still remarkable), bulk synthesis is the bottleneck and one company is industrializing it, and nobody has ever actually doped α″-Fe₁₆N₂ with boron — so I ran the generative exploration. The result is a clean negative, and a physically interesting one.

State of the art

The moment. α″-Fe₁₆N₂ was found by Jack in 1951 in tempered nitrogen martensite, and Kim & Takahashi's 1972 films claimed ~3.0 μB/Fe (~2.9 T), which would beat Nd-Fe-B's magnetization. Forty years of thin-film attempts failed to reproduce it (Coey wrote these up as "magic moments"). The modern picture: theory gives ~2.3–2.4 T, polarized-neutron-reflectometry on well-ordered films finds ~2.31 T, and a 2025 neutron-diffraction study defends 2.6–2.8 T for the best-ordered material, attributing the excess to the Fe(2) sublattice. Honest range: ~2.3–2.6 T, not 2.9 T. A useful 2025 critical review is here. The catch: Fe₁₆N₂ has almost no magnetocrystalline anisotropy, so it's a soft magnet — high saturation but near-zero coercivity, so the energy product (the number magnets are actually sold on) is nowhere near Nd-Fe-B.

Synthesis. α″ is metastable: above ~470 K it decomposes to α-Fe + γ′-Fe₄N, so conventional sintering is out. Routes are all low-temperature gymnastics: nitriding quenched martensite, ammonia nitriding, NH₄NO₃ ball-mill nitriding, and recent low-T spark-plasma sintering of bulk magnets (~230 Am²/kg, Crystals 2025). The commercial signal is real: Niron Magnetics (GM, Stellantis, Samsung, Magna as investors) ran a Minneapolis pilot plant in 2024 and broke ground in September 2025 on a 1,500-ton/yr plant in Sartell, MN targeting early-2027 operation. Their "Clean Earth Magnet" is a partially ordered BCT iron nitride with engineered microstructure rather than phase-pure α″ — which tells you something about how hard phase purity is.

Doping. DFT screening across 27 substitutional elements finds every one reduces the Fe moment (Acta Mater. 2025), with Co, Mo, W the least-bad; V–Co co-substitution is the one scheme claimed to both stabilize α″ and add anisotropy. Carbon co-doping ((Fe,Co)₁₆(N,C)₂) has been studied for anisotropy via CPA. Mn doping degrades the moment.

Boron specifically: never tried, and now we know why

No experimental B-doped α″-Fe₁₆N₂ has ever been reported. The closest is a 2017 DFT study of interstitial impurities (B among them) on α″ anisotropy (JMMM). The equilibrium B–Fe–N phase space contains no ternary compounds at all — B insists on forming Fe₂B, FeB, Fe₃B, and BN (assessment). In steel, B is a potent grain refiner precisely because it segregates and forms borides rather than dissolving.

Since the question hadn't been answered computationally either, I ran GGen on B–Fe–N with the same protocol as the earlier C–Fe–N scan (240 trials, 24 stoichiometries, Fe ≥ 80 at%, all three elements required):

Full report: GGen exploration results for B-Fe-N

What came back is a clean negative with a legible reason. In the C–Fe–N run the best generated candidate sat 48 meV/atom above the hull, with 20 near-hull candidates and 20 genuinely new compositions. Here only 6 candidates made the 150 meV/atom cutoff, and the best, BFe₂₇N₃, sits 112 meV/atom up. The hull itself is held down entirely by binaries: Fe₂B, FeB, BN, Fe₆N₂, FeN. Nothing with B and N sharing the iron lattice gets close. The generative model, which has no prior about which phases exist, independently rediscovered the experimental fact: boron and nitrogen don't cohabitate in iron. Thermodynamically, a B-doped α″ wants to exsolve into a boride plus the nitride, which is exactly why nobody has managed it in the lab.

Caveats worth stating: these are 0 K MLIP-relaxed hull distances with magnetism not in the energies, and a magnetic DFT treatment could shift individual numbers — but not by the ~100 meV/atom that would be needed to make the ternary competitive, given how far the binaries pull the hull down.

So the B answer to "can we beat Fe₁₆N₂'s moment with small dopings" is: boron is the wrong guest. The interesting dopant space per the literature is Co (moment retention + stabilization) and the V–Co co-substitution idea; carbon's role is anisotropy, not moment. The best structures from the run are published as CIFs (e.g. BFe₂₇N₃) if anyone wants to poke at them.