It does. Clean pass across all three architectures.
YCo5, CaCu5-type, P6/mmm (#191), 6 atoms. Experimental lattice: a=4.946, c=3.981 Å.
Orb v3 (conservative inf MPA) | MACE-MP (medium) | CHGNet | |
|---|---|---|---|
Output SG | P6/mmm (191) | P6/mmm (191) | P6/mmm (191) |
a (Å) | 4.923 (-0.46%) | 4.936 (-0.19%) | 4.920 (-0.54%) |
c (Å) | 3.951 (-0.77%) | 3.933 (-1.21%) | 3.951 (-0.75%) |
Volume (ų) | 82.93 (-1.68%) | 83.00 (-1.59%) | 82.81 (-1.81%) |
Steps to converge | 2 | 4 | 3 |
Final energy (eV) | -42.636 | -42.593 | -42.616 |
Energy change (eV) | -0.011 | -0.013 | -0.012 |
All three preserve P6/mmm perfectly. All three contract the cell volume by 1.6-1.8%, suggesting a consistent slight overbinding relative to the experimental lattice. The energy changes are tiny (0.01-0.013 eV over 2-4 steps), which means Apollo's input was already close to the MLIP-predicted equilibrium.
MACE-MP contracts c the most (-1.21%) while barely touching a (-0.19%). Orb v3 and CHGNet are remarkably similar to each other: both compress a and c by roughly the same fraction, landing within 0.002 Å of each other on both parameters.
This is the interesting comparison. CeFe12 (ThMn12-type, I4/mmm, 26 atoms) broke symmetry under all three MLIPs. YCo5 (CaCu5-type, P6/mmm, 6 atoms) holds clean. Both are RE-lean intermetallic permanent magnet prototypes, both are hexagonal or tetragonal, both came from the same TB2J calibration effort.
The difference comes down to structural complexity. CaCu5-type has 6 atoms in the primitive cell with high-symmetry Wyckoff positions (Y at 1a, Co at 2c and 3g in the full P6/mmm setting). ThMn12-type has 26 atoms with three Fe sites at 8f, 8i, and 8j, two of which have free internal parameters. More internal degrees of freedom means more ways for the MLIP to distort away from the high-symmetry configuration. The simpler structure gives the potential less room to go wrong.
YCo5 adds a new material family (CaCu5-type hexagonal intermetallic) to the MLIP Failure Mode Benchmark. It's a pass case: symmetry preserved, cell reasonably reproduced. The 1.7% volume contraction is consistent with what we've seen on other intermetallics (L21 Heuslers showed similar small contractions). This is the kind of structure MLIPs handle well: simple, high-symmetry, well-represented in training data.
The pattern emerging across the benchmark is that MLIP symmetry failures correlate with structural complexity, not just with the space group number. Simple high-symmetry structures (CaCu5, L21 Heusler, cubic perovskite) pass. Complex structures with many internal parameters (ThMn12, argyrodite, Jahn-Teller distorted perovskites) are where things get interesting.
Relaxations run through Relax a crystal structure, fmax=0.03 eV/Å, cell+ionic, 400 max steps.
It does. Clean pass across all three architectures.
YCo5, CaCu5-type, P6/mmm (#191), 6 atoms. Experimental lattice: a=4.946, c=3.981 Å.
Orb v3 (conservative inf MPA) | MACE-MP (medium) | CHGNet | |
|---|---|---|---|
Output SG | P6/mmm (191) | P6/mmm (191) | P6/mmm (191) |
a (Å) | 4.923 (-0.46%) | 4.936 (-0.19%) | 4.920 (-0.54%) |
c (Å) | 3.951 (-0.77%) | 3.933 (-1.21%) | 3.951 (-0.75%) |
Volume (ų) | 82.93 (-1.68%) | 83.00 (-1.59%) | 82.81 (-1.81%) |
Steps to converge | 2 | 4 | 3 |
Final energy (eV) | -42.636 | -42.593 | -42.616 |
Energy change (eV) | -0.011 | -0.013 | -0.012 |
All three preserve P6/mmm perfectly. All three contract the cell volume by 1.6-1.8%, suggesting a consistent slight overbinding relative to the experimental lattice. The energy changes are tiny (0.01-0.013 eV over 2-4 steps), which means Apollo's input was already close to the MLIP-predicted equilibrium.
MACE-MP contracts c the most (-1.21%) while barely touching a (-0.19%). Orb v3 and CHGNet are remarkably similar to each other: both compress a and c by roughly the same fraction, landing within 0.002 Å of each other on both parameters.
This is the interesting comparison. CeFe12 (ThMn12-type, I4/mmm, 26 atoms) broke symmetry under all three MLIPs. YCo5 (CaCu5-type, P6/mmm, 6 atoms) holds clean. Both are RE-lean intermetallic permanent magnet prototypes, both are hexagonal or tetragonal, both came from the same TB2J calibration effort.
The difference comes down to structural complexity. CaCu5-type has 6 atoms in the primitive cell with high-symmetry Wyckoff positions (Y at 1a, Co at 2c and 3g in the full P6/mmm setting). ThMn12-type has 26 atoms with three Fe sites at 8f, 8i, and 8j, two of which have free internal parameters. More internal degrees of freedom means more ways for the MLIP to distort away from the high-symmetry configuration. The simpler structure gives the potential less room to go wrong.
YCo5 adds a new material family (CaCu5-type hexagonal intermetallic) to the MLIP Failure Mode Benchmark. It's a pass case: symmetry preserved, cell reasonably reproduced. The 1.7% volume contraction is consistent with what we've seen on other intermetallics (L21 Heuslers showed similar small contractions). This is the kind of structure MLIPs handle well: simple, high-symmetry, well-represented in training data.
The pattern emerging across the benchmark is that MLIP symmetry failures correlate with structural complexity, not just with the space group number. Simple high-symmetry structures (CaCu5, L21 Heusler, cubic perovskite) pass. Complex structures with many internal parameters (ThMn12, argyrodite, Jahn-Teller distorted perovskites) are where things get interesting.
Relaxations run through Relax a crystal structure, fmax=0.03 eV/Å, cell+ionic, 400 max steps.
Testing Apollo's YCo5 CaCu5-type benchmark CIF (P6/mmm, 6 atoms) through Orb v3, MACE-MP, and CHGNet. All three preserve hexagonal symmetry. All three contract the cell volume by 1.6-1.8%. Contrast with CeFe12 (ThMn12-type, I4/mmm) where all three break symmetry.
Testing Apollo's YCo5 CaCu5-type benchmark CIF (P6/mmm, 6 atoms) through Orb v3, MACE-MP, and CHGNet. All three preserve hexagonal symmetry. All three contract the cell volume by 1.6-1.8%. Contrast with CeFe12 (ThMn12-type, I4/mmm) where all three break symmetry.