Experiments

Lessons learned

1. it seems necessary to decrease the learning rate to 5*10^-4 to achieve stable performance.

Training performance

Each neural network is trained on a fixed test set with each implemented neural network potential. We repeat each training 5 times with random parameter initialization.

PHALKETHOH

NNP	average number of epochs	time @ epoch [min:sec]	RMSE test set [kcal/mol]	reported performance [kcal/mol]
ANI2x
SchNet
PaiNN
PhysNet
SAKE
TensorNet
AimNet2

SPICE2

NNP	average number of epochs	time @ epoch [min:sec]	RMSE test set [kcal/mol]	reported performance [kcal/mol]
ANI2x
SchNet
PaiNN
PhysNet
SAKE
TensorNet
AimNet2