Table 2 Energy and Force RMSE for the 3BPA temperature transferability dataset, reported in units of [meV] and [meV/Å]

From: Learning local equivariant representations for large-scale atomistic dynamics

 

ACE12

sGDML10

GAP6

FF46, 47

ANI-pretrained48, 49

ANI-2x48, 49

NequIP15

Allegro

Fit to 300 K

        

300 K, E

7.1

9.1

22.8

60.8

23.5

38.6

3.28 (0.12)

3.84 (0.10)

300 K, F

27.1

46.2

87.3

302.8

42.8

84.4

10.77 (0.28)

12.98 (0.20)

600 K, E

24.0

484.8

61.4

136.8

37.8

54.5

11.16 (0.17)

12.07 (0.55)

600 K, F

64.3

439.2

151.9

407.9

71.7

102.8

26.37 (0.11)

29.11 (0.27)

1200 K, E

85.3

774.5

166.8

325.5

76.8

88.8

38.52 (2.00)

42.57 (1.79)

1200 K, F

187.0

711.1

305.5

670.9

129.6

139.6

76.18 (1.36)

82.96 (2.17)

  1. All models were trained on T = 300 K. Results for all models except for NequIP and Allegro from ref. 24. Best results are marked in bold. For NequIP and Allegro, we report the mean over three different seeds as well as the sample standard deviation in parentheses.
Back to article page