Table 1 Performance of four versions of the CACE-LR potentials on each test set

From: Machine learning interatomic potential can infer electrical response

Version

1

2

3

4

 

E+F

E+F

E+F

E+F+BEC

Nconfig

654

654

100

100

rcut(Ă…)

5.5

4.5

4.5

4.5

E

0.22

0.25

0.26

0.19

F

18.88

21.01

23.84

25.34

\({Z}_{\alpha \alpha }^{* }\)

0.11

0.06

0.05

0.04

\({Z}_{\alpha \beta }^{* }\)

0.04

0.04

0.05

0.03

\({Z}_{\alpha \alpha }^{* }\,{R}^{2}\)

0.97

0.99

0.99

1.00

\({Z}_{\alpha \beta }^{* }\,{R}^{2}\)

0.90

0.90

0.89

0.94

  1. Reported metrics are test root mean squared errors (RMSE)s: in meV/atom for energy (E), meV/ Å for forces (F), and e for Born effective charge (BEC) tensors, separated into diagonal (\({Z}_{\alpha \alpha }^{* }\)) and off-diagonal (\({Z}_{\alpha \beta }^{* }\) with α ≠ β) components. R2 coefficients for BEC components are also included.