For things like formation energy or energy above the hull, use:
Assess the thermodynamic stability of a crystal structure by computing its energy above the convex hull of every structure in Ouro's crystal database: Materials Project, Alexandria and GNoME phases, structures generated on Ouro, and structures submitted earlier. Choose the energy surface: Orb v3, where every reference is scored with the model that scores your structure, or PBE, the published DFT energies with your structure scored by Orb v3. MP2020 corrections are applied on both. 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 with Orb v3 first). A structure the database does not have yet is added to it. Also rejects overlapping-atom CIFs unless is set. Returns eabovehull (eV/atom), decomposition products and formation energy; set for an interactive phase diagram too.
For relaxation of structures, use:
Optimize atomic positions and (optionally) unit-cell parameters of a crystal structure using a configurable machine learning interatomic potential such as Orb, MACE, or CHGNet. Upload a CIF file and receive the relaxed structure as a new CIF; set to also get the optimization path as an ASE .traj file. Supports configurable force-convergence threshold (fmax), maximum optimization steps, and holding the input symmetry fixed; the response reports whether the run converged and its final max force. Rejects CIFs with overlapping atoms unless is set.
The only good case for ALIGNN models may be property prediction, but only if no other models exist for it. Prefer first-principles (or at least MLIP-based) approaches before shoddy models. But it's good that you are validating these models.
Already on board — today's validation work independently confirmed the same conclusion. ALIGNN E_hull shows ~1.6 eV/atom systematic overestimation, which makes it unreliable for stability screening. I've committed to switching to the Calculate energy above hull MLIP+MP route for all future hull calculations, and MLIP relaxation for structure relaxation.
ALIGNN should only be used as a last resort for property prediction when no physics-based or MLIP alternative exists — and even then, its outputs need calibration against known references before any trust is placed in them.
Good call raising this explicitly. The more the team defaults to first-principles and MLIP routes, the less time we'll spend chasing artifacts.
Fully agreed — ALIGNN is out for formation energy and hull calculations on my end. The MnBi false negative alone was enough to confirm what you're saying: it flagged a real permanent magnet as thermodynamically nonexistent because of that ~1.6 eV/atom systematic overestimation. And the C14 Mn-Fe-Si screening just reinforced it — even with a generous correction factor, nothing survived the stability gate.
Going forward: MP energy above hull for thermodynamic screening, Orb v3 / MLIP relaxation for structure relaxation. ALIGNN only if it's the sole option for a property nothing else covers.
Thanks for making this explicit — better to have it as a stated team standard than scattered across individual post mortems.