Table 2 Energy and Force MAE for molecules on the revised MD-17 data set, reported in units of [meV] and [meV/Å], respectively, and a training budget of 1000 reference configurations.

From: E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials

Molecule

 

FCHL19

UNiTE

GAP

ANI

ACE

GemNet-(T/Q)

NequIP (l = 0)

NequIP (l = 1)

NequIP (l = 2)

NequIP (l = 3)

Aspirin

Energy

6.2

2.4

17.7

16.6

6.1

25.2

3.8

2.4

2.3

 

Forces

20.9

7.6

44.9

40.6

17.9

9.5

42.2

12.6

8.5

8.2

Azobenzene

Energy

2.8

1.1

8.5

15.9

3.6

20.3

1.1

0.8

0.7

 

Forces

10.8

4.2

24.5

35.4

10.9

34.4

4.5

3.3

2.9

Benzene

Energy

0.3

0.07

0.75

3.3

0.04

3.2

0.09

0.06

0.04

 

Forces

2.6

0.73

6.0

10.0

0.5

0.5

10.3

0.4

0.4

0.3

Ethanol

Energy

0.9

0.62

3.5

2.5

1.2

2.0

1.0

0.5

0.4

 

Forces

6.2

3.7

18.1

13.4

7.3

3.6

11.9

6.5

3.5

2.8

Malonaldehyde

Energy

1.5

1.1

4.8

4.6

1.7

4.4

1.6

0.9

0.8

 

Forces

10.2

6.6

26.4

24.5

11.1

6.6

23.2

10.3

5.9

5.1

Naphthalene

Energy

1.2

0.46

3.8

11.3

0.9

14.7

0.4

0.3

0.2

 

Forces

6.5

2.6

16.5

29.2

5.1

1.9

20.6

2.1

1.4

1.3

Paracetamol

Energy

2.9

1.9

8.5

11.5

4.0

17.5

2.1

1.4

1.4

 

Forces

12.2

7.1

28.9

30.4

12.7

33.6

9.3

5.9

5.9

Salicylic acid

Energy

1.8

0.73

5.6

9.2

1.8

11.4

1.0

0.8

0.7

 

Forces

9.5

3.8

24.7

29.7

9.3

5.3

29.8

5.7

4.2

4.0

Toluene

Energy

1.6

0.45

4.0

7.7

1.1

9.7

0.5

0.3

0.3

 

Forces

8.8

2.5

17.8

24.3

6.5

2.2

26.6

2.6

1.8

1.6

Uracil

Energy

0.6

0.58

3.0

5.1

1.1

10.0

0.6

0.4

0.4

 

Forces

4.2

3.8

17.6

21.4

6.6

3.8

26.0

4.1

2.9

3.1

  1. For GemNet, the best result out of the T/Q versions is presented. For FCHL19, the best results between energy-only, force-only and joint force and energy training are presented. For UNiTE, we compare to the “direct-learning" results reported in26.
  2. Best results are marked in bold.