Can a model that also predicts magnetism replace the main energy model?

EXPERIMENTS.md · lines 875–915

E20 — CHGNet: worse energies, but it knows when a system is magnetic

Date 2026-09-12 · Question Is there a surrogate that handles spin, and should it replace or join MACE-MPA-0? · Provenance …/chgnet_compare.py

Why it was asked. MACE-MP-0 and MACE-MPA-0 do not predict spin. Nickel superalloys involve Ni, Co, Fe and Cr, all magnetic, so extending the environment there raises the question of whether a charge- or spin-aware model is needed. CHGNet predicts magnetic moments per atom as a fourth output alongside energy, forces and stress, having been trained on MPtrj's magmom labels.

Method. CHGNet 0.3.0 (via chgnet 0.4.2) on the same 200 near-equilibrium (|strain| <= 5%) RHEA bcc solid solutions used in E18, per-element referencing as before.

Result.

meV/atom. CHGNet is four times worse on energy and about ten times faster, which is what 22x fewer parameters buys.

The magnetism sanity check passes. On these non-magnetic refractories CHGNet predicts mean |magmom| 0.069 and max 0.217 uB/atom - correctly near zero - while returning 2.51 uB/atom on bcc Fe against a PBE and experimental value near 2.2. It distinguishes magnetic from non-magnetic systems.

Interpretation. CHGNet should not replace MACE-MPA-0 for energies. Its useful role is as a trust flag: run its magmom prediction on each candidate, and where it returns significant spin, the candidate sits in a regime a non-spin-aware model cannot represent, so it should be marked untrusted rather than silently screened. That does not fix the energy, but it marks the boundary of the envelope, which is what E19 showed an environment needs.

Not tested. Neither model has been evaluated on actually magnetic alloys - RHEA contains no Fe, Co or Ni. CHGNet's energy accuracy on ferromagnetic systems, where its charge-informed training is supposed to pay off, is unmeasured here and is the thing that would decide its role for superalloys.