Table 3 Summary of accuracy for all trained models (cutoff 0.6 nm) on the OE62 dataset compared to other long-range modeling approaches

From: Learning non-local molecular interactions via equivariant local representations and charge equilibration

Model

U MAE [meV]

U RMSE [meV]

# Mio. Params.

Allegro (–)

Baseline S

63.4

123.7

0.17

Baseline S+

60.0

114.5

0.20

Baseline L

61.1

120.9

0.19

Baseline L+

61.8

116.6

0.22

CELLI S

55.3

116.7

0.21

CELLI L

55.1

114.3

0.29

MACE (2)

Baseline

48.1

90.1

2.37

CELLI

48.0

88.3

2.52

DimeNet++ (3)

Baseline

42.1 (53.8*)

108.4

2.78

Ewald12

48.1

–

4.8

Neural P3M21

41.5

–

–

PaiNN (4)

Ewald12

59.7

–

15.7

Neural P3M21

52.9

–

–

  1. The Small (S) versions of CELLI and Allegro were used to compute the benchmarks and use fewer irreps, a lower rotational order for the spherical harmonics, and a smaller hidden size for the charge-embedding networks than the Large (L) version. The versions S+ and L+ of Allegro contain one additional Interaction Layer compared to the CELLI version. The number of message-passing steps of each model, if applicable, is reported in brackets behind the model name. Results for Ewald and Neural P3M on DimeNet++ and PaiNN were taken directly from the refs. 12,21. The lowest errors are reported in bold. (*) Reported by Kosmala et al.
Back to article page