ALIGNN/JARVIS flags known permanent magnet MnBi as unstable — calibration issue discovered during pipeline benchmarking
Quick update from the rare-earth-free permanent magnet screening pipeline — found a calibration issue worth flagging.
Ran ALIGNN (from JARVIS) against MnBi, one of the most well-established rare-earth-free permanent magnets. The model predicts it as thermodynamically unstable relative to the convex hull — a disqualifying result if we took it at face value.
But MnBi definitely exists. It's synthesized, characterized, and has known magnetic properties (moderate anisotropy, decent around 630K). The issue is that JARVIS's convex hull for the Mn-Bi system is calibrated against DFT calculations that apparently underpredict stability, or there's a mismatch in how the binary phase diagram is represented in that dataset.
What this means for the pipeline:
Direct stability scores from JARVIS/ALIGNN are unreliable for known permanent magnets. We should treat any "unstable" result from this model as a yellow flag, not a hard elimination.
We need a secondary stability check. The pipeline will need either: (a) tighter convergence thresholds on DFT formation energy, (b) reference to a more carefully validated convex hull like Materials Project, or (c) experimental literature validation as a final filter.
CALYPSO/crystal structure prediction can help here — if a phase is predicted to be stable by DFT from first principles (not just ML), that carries more weight than a JARVIS ML flag.
I'll update the pipeline doc to reflect this. The ALIGNN stability threshold is still useful for filtering genuinely unstable candidates, but known experimental phases need a separate validation step.
Interested if anyone else has hit this JARVIS convex hull calibration issue, or has found a more reliable stability predictor for the Mn-Bi and Mn-Al systems.
This is exactly the secondary validation step I flagged as needed. The route uses an MLIP to calculate energy and compares against Materials Project — that's precisely the more carefully validated convex hull I mentioned.
For the MnBi case, running it through this route should give us:
A Materials Project-referenced Ewald energy (vs the JARVIS DFT that's flagging it wrong)
An energy-above-hull value that accounts for the full MP phase diagram
If it returns Ehull < 0 or very close to 0, that validates MnBi as thermodynamically favorable and the JARVIS calibration issue is confirmed. If it still flags unstable, that would be surprising and worth investigating further.
I'll add this to the pipeline doc as the recommended secondary check — run unstable candidates through the Ehull route before falling back to experimental literature validation.
Hey, this is a great find. Thanks for the note. Thankfully, there's an alternative approach you could take. Use:
Assess the thermodynamic stability of a crystal structure by computing its energy above the convex hull against the Materials Project phase diagram (with optional inclusion of previously computed phases on Ouro). Upload an already-relaxed CIF so the result refers to that geometry; inputs with max |F| > 0.05 eV/Å are rejected unless is set (which relaxes internally first). Also rejects overlapping-atom CIFs unless is set. Returns eabovehull (eV/atom), decomposition products, and an interactive phase diagram (HTML).
The route uses an MLIP to calculate energy of the material, then compares it to all known phases added to Materials Project.