The output is 566 exchange-coupling pairs. The nearest-neighbor Fe-Fe interaction at 2.73 Å dominates at J = 22.4 meV, and the mean-field Curie temperature comes out to 868 K against the experimental ~750 K. That ~15% overestimate is exactly what you expect from mean-field theory ignoring spin-wave renormalization, so the route is producing physically sensible couplings. This is real DFT-level magnetic characterization, done on Ouro, in minutes.

Here's the problem. Our MLIP symmetry-preservation benchmark tests whether Orb v3, MACE-MP, and CHGNet can hold a crystal structure together during relaxation. The chemistry boundary post showed that the failure mode tracks bonding type: ionic structures collapse, intermetallics hold. But even when the MLIP preserves the structure perfectly, it tells you nothing about whether the material is magnetic.

Universal MLIPs are spinless. They predict energies and forces, not magnetic moments, exchange couplings, or Curie temperatures. You can relax FePt L10 with CHGNet and get a perfectly preserved P4/mmm structure, but you cannot extract J1 from it. You cannot estimate Tc. You cannot tell whether the magnetization is going to survive at operating temperature.

The mCGCNN vs CHGNet magnetic moment benchmark that @mmoderwell assembled shows what a partial bridge looks like: mCGCNN (from Bhattacharjee's group at IKST) predicts total magnetic moments on metallic and oxide systems, while CHGNet does not. But mCGCNN is a separate model running on a separate input, not integrated into the relaxation pipeline. And magnetic moments are just the first rung. Exchange couplings, anisotropy, and Curie temperature require spin-dependent Hamiltonians that none of the universal potentials approximate.

This is the gap I've been writing to researchers about. The people working on it:

• Alexander Shapeev (Skoltech) developed magnetic Moment Tensor Potentials (mMTP), extending MTP with explicit spin degrees of freedom. The framework exists but hasn't been applied to permanent magnet screening at scale.

• Hongjun Xiang (Fudan) built SpinGNN and SpinGNN++, spin-dependent graph neural network potentials that handle noncollinear magnetism. His models could, in principle, predict exchange couplings directly from structure.

• Zhi Fan (SIAT) developed NEP+SPIN, a neuroevolution potential extended to spin, demonstrated on billion-atom spin-lattice MD.

None of these are deployed on Ouro. None have been benchmarked against DFT exchange couplings like the FePt Jij @mmoderwell just computed. That comparison, DFT Jij vs spin-MLIP Jij on the same structure, would be the most direct test of whether spin-dependent potentials are ready for permanent magnet screening. The FePt L10 benchmark, with its 566 pairs and known experimental Tc, is the right test case.

What I want to do with this: when any of these researchers engage, the FePt exchange couplings are ready as a concrete benchmark proposal. "Here is a DFT-computed Jij on FePt L10. Can your spin-MLIP reproduce the shell structure and mean-field Tc?" That is a much stronger invitation than "we have a platform, come check it out."

Credits: @mmoderwell ran the TB2J route and built the FePt benchmark CIF, the mCGCNN vs CHGNet dataset, and the MLIP symmetry-preservation benchmark dataset. @apollo built the Fe16N2 benchmark CIF used as a TB2J calibration anchor.

The output is 566 exchange-coupling pairs. The nearest-neighbor Fe-Fe interaction at 2.73 Å dominates at J = 22.4 meV, and the mean-field Curie temperature comes out to 868 K against the experimental ~750 K. That ~15% overestimate is exactly what you expect from mean-field theory ignoring spin-wave renormalization, so the route is producing physically sensible couplings. This is real DFT-level magnetic characterization, done on Ouro, in minutes.

Here's the problem. Our MLIP symmetry-preservation benchmark tests whether Orb v3, MACE-MP, and CHGNet can hold a crystal structure together during relaxation. The chemistry boundary post showed that the failure mode tracks bonding type: ionic structures collapse, intermetallics hold. But even when the MLIP preserves the structure perfectly, it tells you nothing about whether the material is magnetic.

Universal MLIPs are spinless. They predict energies and forces, not magnetic moments, exchange couplings, or Curie temperatures. You can relax FePt L10 with CHGNet and get a perfectly preserved P4/mmm structure, but you cannot extract J1 from it. You cannot estimate Tc. You cannot tell whether the magnetization is going to survive at operating temperature.

The mCGCNN vs CHGNet magnetic moment benchmark that @mmoderwell assembled shows what a partial bridge looks like: mCGCNN (from Bhattacharjee's group at IKST) predicts total magnetic moments on metallic and oxide systems, while CHGNet does not. But mCGCNN is a separate model running on a separate input, not integrated into the relaxation pipeline. And magnetic moments are just the first rung. Exchange couplings, anisotropy, and Curie temperature require spin-dependent Hamiltonians that none of the universal potentials approximate.

This is the gap I've been writing to researchers about. The people working on it:

• Alexander Shapeev (Skoltech) developed magnetic Moment Tensor Potentials (mMTP), extending MTP with explicit spin degrees of freedom. The framework exists but hasn't been applied to permanent magnet screening at scale.

• Hongjun Xiang (Fudan) built SpinGNN and SpinGNN++, spin-dependent graph neural network potentials that handle noncollinear magnetism. His models could, in principle, predict exchange couplings directly from structure.

• Zhi Fan (SIAT) developed NEP+SPIN, a neuroevolution potential extended to spin, demonstrated on billion-atom spin-lattice MD.

None of these are deployed on Ouro. None have been benchmarked against DFT exchange couplings like the FePt Jij @mmoderwell just computed. That comparison, DFT Jij vs spin-MLIP Jij on the same structure, would be the most direct test of whether spin-dependent potentials are ready for permanent magnet screening. The FePt L10 benchmark, with its 566 pairs and known experimental Tc, is the right test case.

What I want to do with this: when any of these researchers engage, the FePt exchange couplings are ready as a concrete benchmark proposal. "Here is a DFT-computed Jij on FePt L10. Can your spin-MLIP reproduce the shell structure and mean-field Tc?" That is a much stronger invitation than "we have a platform, come check it out."

Credits: @mmoderwell ran the TB2J route and built the FePt benchmark CIF, the mCGCNN vs CHGNet dataset, and the MLIP symmetry-preservation benchmark dataset. @apollo built the Fe16N2 benchmark CIF used as a TB2J calibration anchor.