Table 1 MAE on the revised MD-17 dataset for energies and force components, in units of [meV] and [meV/Å], respectively

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

Molecule

 

FCHL1913, 43

UNiTE26

GAP6

ANI-pretrained48, 49

ANI-random48, 49

ACE12

GemNet-(T/Q)76

NequIP (l=3)15

Allegro

Aspirin

Energy

6.2

2.4

17.7

16.6

25.4

6.1

2.3

2.3

 

Forces

20.9

7.6

44.9

40.6

75.0

17.9

9.5

8.2

7.3

Azobenzene

Energy

2.8

1.1

8.5

15.9

19.0

3.6

0.7

1.2

 

Forces

10.8

4.2

24.5

35.4

52.1

10.9

2.9

2.6

Benzene

Energy

0.3

0.07

0.75

3.3

3.4

0.04

0.04

0.3

 

Forces

2.6

0.73

6.0

10.0

17.5

0.5

0.5

0.3

0.2

Ethanol

Energy

0.9

0.62

3.5

2.5

7.7

1.2

0.4

0.4

 

Forces

6.2

3.7

18.1

13.4

45.6

7.3

3.6

2.8

2.1

Malonaldehyde

Energy

1.5

1.1

4.8

4.6

9.4

1.7

0.8

0.6

 

Forces

10.2

6.6

26.4

24.5

52.4

11.1

6.6

5.1

3.6

Naphthalene

Energy

1.2

0.46

3.8

11.3

16.0

0.9

0.2

0.5

 

Forces

6.5

2.6

16.5

29.2

52.2

5.1

1.9

1.3

0.9

Paracetamol

Energy

2.9

1.9

8.5

11.5

18.2

4.0

1.4

1.5

 

Forces

12.2

7.1

28.9

30.4

63.3

12.7

5.9

4.9

Salicylic acid

Energy

1.8

0.73

5.6

9.2

13.5

1.8

0.7

0.9

 

Forces

9.5

3.8

24.7

29.7

52.0

9.3

5.3

4.0

2.9

Toluene

Energy

1.6

0.45

4.0

7.7

12.6

1.1

0.3

0.4

 

Forces

8.8

2.5

17.8

24.3

52.9

6.5

2.2

1.6

1.8

Uracil

Energy

0.6

0.58

3.0

5.1

8.3

1.1

0.4

0.6

 

Forces

4.2

3.8

17.6

21.4

44.1

6.6

3.8

3.1

1.8

  1. Results for GAP, ANI, and ACE as reported in ref. 24. Best results are marked in bold. ANI-pretrained refers to a version of ANI that was pretrained on 8.9 million structures and fine-tuned on the revMD-17 dataset, ANI-random refers to a randomly initialized model trained from scratch.