Table 1 Performance of MACE-MP-f4 10%, ACE-f 10%, and from-scratch (f-s) ACE and MACE models

From: Fine-tuning foundation models of materials interatomic potentials with frozen transfer learning

Property

MACE-MP-f

ACE-f

ACE

MACE

 

10%

10%

(f-s)

(f-s)

MAE (E)

6.7

4.8

8.6

11.3

RMSE (E)

13.5

6.4

13.7

15.8

MAE (F)

3.1

7.2

10.4

8.1

RMSE (F)

10.1

10.9

18.2

12.9

MAE (H-atom F)

4.3

9.0

28.5

17.7

RMSE (H-atom F)

17.1

17.7

47.6

32.4

teval (E)

372.6

9.6

9.1

60.0

teval (F)

390.5

22.6

20.7

60.0

  1. Energy (E) and force (F) errors are included in meV and meV/Å, respectively. Evaluation times (teval) are in ms and were calculated on a single AMD EPYC 7742 (Rome) 2.25 GHz CPU processor core.