Table 5.
Tensor rank l, feature size, radial cutoff in units of [Å], as well as energy and force weights used in the joint loss function.
| Data Set | Tensor rank l | # Features | rc | λE | λF |
|---|---|---|---|---|---|
| MD-17 | 3 | 64 | 4.0 | 1 | 1000 |
| revMD-17 | {0, 1, 2, 3} | 64 | 4.0 | 1 | 1000 |
| CCSD/CCSD(T) | 3 | 64 | 4.0 | 1 | 1000 |
| Water+Ices, DeepMD | 2 | 32 | 6.0 | see 3 | see 3 |
| Formate on Cu | 2 | 32 | 5.0 | 1 | 2704 |
| Li4P2O7 | 2 | 32 | 5.0. | 1 | 43,264 |
| LiPS | 2 | 32 | 5.0 | 1 | 6889 |
| Water, Cheng et al. | {0, 1, 2, 3} | 32 | 4.5 | 1 | 36,864 |
All models were trained with even and odd features, i.e. a tensor rank of l = 1 and 32 features corresponds to 32x0o + 32x0e + 32x1o + 32x1e. The force weightings for formate on Cu, LiPO, LiPS, and the water system for sample efficiency tests stem from .