Skip to main content
. 2022 May 4;13:2453. doi: 10.1038/s41467-022-29939-5

Table 1.

Energy and Force MAE for molecules on the original MD-17 data set, reported in units of [meV] and [meV/Å], respectively, and a training budget of 1000 reference configurations.

Molecule SchNet DimeNet sGDML PaiNN SpookyNet GemNet-(T/Q) NewtonNet UNiTE NequIP (l = 3)
Aspirin Energy 16.0 8.8 8.2 6.9 6.5 7.3 5.7
Forces 58.5 21.6 29.5 14.7 11.2 9.4 15.1 6.8 8.0
Ethanol Energy 3.5 2.8 3.0 2.7 2.3 2.6 2.2
Forces 16.9 10.0 14.3 9.7 4.1 3.7 9.1 4.0 3.1
Malonaldehyde Energy 5.6 4.5 4.3 3.9 3.4 4.2 3.3
Forces 28.6 16.6 17.8 13.8 7.2 6.7 14.0 6.9 5.6
Naphthalene Energy 6.9 5.3 5.2 5.0 5.0 5.1 4.9
Forces 25.2 9.3 4.8 3.3 3.9 2.2 3.6 2.8 1.7
Salicylic acid Energy 8.7 5.8 5.2 4.9 4.9 5.0 4.6
Forces 36.9 16.2 12.1 8.5 7.8 5.4 8.5 4.2 3.9
Toluene Energy 5.2 4.4 4.3 4.1 4.1 4.1 4.0
Forces 24.7 9.4 6.1 4.1 3.8 2.6 3.8 3.1 2.0
Uracil Energy 6.1 5.0 4.8 4.5 4.6 4.6 4.5
Forces 24.3 13.1 10.4 6.0 5.2 4.2 6.5 4.2 3.3

For GemNet, the best result out of the T/Q versions is presented and for PaiNN the best between force-only and joint force and energy training. For UNiTE, we compare to the “direct-learning” results reported in26.

Best results are marked in bold.