Table 1 Performance comparison between direct training on DFT data without pretraining and transfer learning with pretraining

From: Neural network potentials with effective charge separation for non-equilibrium dynamics of ionic solids: a ZnO case study

 

DATA

E (meV/atom)

F (meV/Å)

σ (MPa)

Direct Training

MOMT (Training)

4.05

83.2

113.9

Transfer Learning

4.11

69.9

113.9

Direct Training

Tens (Unseen)

6.84

73.6

211.7

Transfer Learning

6.66

64.3

176.7

  1. For fair comparison, we utilized the same number of epochs for the initial training (400) and finetuning (1000). We note that even pre-training from empirical potentials, Tens data was not included.
  2. Bold values indicate the lower error for each property (energy, force, and stress) between the two models.
Back to article page