Failure diagnosis (F0–F9)
Each method’s raw, unrelaxed top-1 candidate is classified into a failure-mode bucket, with the frozen MatterSim potential and the campaign matcher. Bars show the share of the 180 targets in each group; sort by any column.
Bucket definitions
| F0 | matches as generated |
| F1 / F2 | matches after loose symmetrisation / after volume rescaling |
| F3a / F3b | matches after symmetry-constrained / free relaxation |
| F8 | no match, but the freely relaxed candidate is more than 5 meV/atom below the relaxed reference |
| F4 / F5 | decorated Wyckoff sites right / space group right |
| F6 / F7 | crystal system or topology right / local coordination right (CN accuracy > 0.9) |
| F9 | wrong local chemistry |
Counts are over the targets for which the method returned a candidate (CrystalFormer: 84). MaskGXT has 176 diagnosed targets (4 unsupported). The F8 bucket is “lower in energy than the reference under MatterSim”; it is not DFT-verified.