Fig. 7. Performance of MLPs trained on energies and forces. Root-mean squared error (RMSE) calculated for energies (top panel) and forces (bottom) with different ML potentials for all molecules of the MD17 database. The models were trained on a sub-sample of 1000 molecular geometries for each system. The reported RMSE values correspond to the average of the test errors for 20 independently trained models evaluated on a test set with 20k geometries.