Table 3.
Comparison of models on the QM9 dataset, measured by the MAE in units of [meV]
| Model | U0 | U | H | G |
|---|---|---|---|---|
| Schnet25 | 14 | 19 | 14 | 14 |
| DimeNet++77 | 6.3 | 6.3 | 6.5 | 7.6 |
| Cormorant23 | 22 | 21 | 21 | 20 |
| LieConv78 | 19 | 19 | 24 | 22 |
| L1Net79 | 13.5 | 13.8 | 14.4 | 14.0 |
| SphereNet80 | 6.3 | 7.3 | 6.4 | 8.0 |
| EGNN40 | 11 | 12 | 12 | 12 |
| ET38 | 6.2 | 6.3 | 6.5 | 7.6 |
| NoisyNodes81 | 7.3 | 7.6 | 7.4 | 8.3 |
| PaiNN27 | 5.9 | 5.7 | 6.0 | 7.4 |
| Allegro, 1 layer | 5.7 (0.3) | 5.3 | 5.3 | 6.6 |
| Allegro, 3 layers | 4.7 (0.2) | 4.4 | 4.4 | 5.7 |
Allegro outperforms all existing atom-centered message-passing and transformer-based models, in particular even with a single layer. Best methods are shown in bold.