Table 1.
Test results of PaiNN models trained on ANI1x, QM9x, and Transition1x.
| Trained on | Tested on | Energy [eV] | Forces [eV/Å] | ||
|---|---|---|---|---|---|
| RMSE | MAE | RMSE | MAE | ||
| ANI1x | Transition States | 0.629 (11) | 0.495 (10) | 0.71(2) | 0.53(1) |
| Transition1x | 0.112 (3) | 0.075 (1) | 0.211(1) | 0.111(1) | |
| QM9x | 3.132 (23) | 2.957 (25) | 0.71(2) | 0.316(5) | |
| ANI1x | ANI1x | 0.044(5) | 0.023(1) | 0.062(1) | 0.039(1) |
| Transition1x | 0.365(17) | 0.226(8) | 0.43(3) | 0.179(1) | |
| QM9x | 3.042(13) | 2.313(11) | 1.9(1) | 1.29(1) | |
| ANI1x | Transition1x | 0.628(63) | 0.289(13) | 0.65(1) | 0.20(8) |
| Transition1x | 0.102(2) | 0.048(1) | 0.136(1) | 0.058(1) | |
| QMx | 2.613(18) | 1.421(11) | 0.495(3) | 0.241(1) | |
| ANI1x | QM9x | 0.134(1) | 0.124(1) | 0.061(1) | 0.038(2) |
| Transition1x | 0.111(2) | 0.074(3) | 0.082(1) | 0.048(1) | |
| QM9x | 0.04(2) | 0.015(1) | 0.016(0) | 0.007(0) | |
We report RMSE and MAE on energy and forces. Force error is the component-wise error between the predicted and true force vector. The test sets have been constructed such that all configurations contain C, N, O, and H, and such that no formula has been seen previously in the training data. We show the best performing model in bold in each test-setup.