Table 4 Comparison of MAE for 12 molecular property prediction tasks on the QM9 dataset

From: Adaptive edge-aware graph convolutional with multi-task learning for simultaneous prediction of material properties

properties

unit

SchNet

MEGNet

Physnet

DimeNet

ET

R-MAT

GeoT

Ours

μ

D

0.033

0.040

0.0529

0.0286

0.011

0.110

0.0297

0.053

α

\({a}_{0}^{3}\)

0.235

0.083

0.0615

0.0469

0.059

0.082

0.052

0.052

\({\varepsilon }_{{HOMO}}\)

meV

41.0

38.0

32.9

27.8

20.3

31

25

23.1

\({\varepsilon }_{{LUMO}}\)

meV

34.0

31.0

24.7

19.7

17.5

29

20.2

22.0

\({\Delta }_{\varepsilon }\)

meV

63.0

61.0

42.5

34.8

36.1

48

43

32.1

R2

\({a}_{0}^{3}\)

0.073

0.265

0.765

0.331

0.033

0.676

0.30

0.832

ZPVE

meV

1.70

1.40

1.39

1.29

1.84

2.23

1.70

1.90

\({U}_{0}\)

meV

14.0

9.0

8.15

8.02

6.16

12.0

11.1

10.3

U

meV

19.0

10.0

8.34

7.89

6.38

10.0

11.7

10.5

H

meV

14.0

10.0

8.42

8.11

6.16

10.0

11.3

10.2

G

meV

14.0

10.0

9.40

8.98

7.62

10.0

11.7

10.7

Cv

\(\frac{{cal}}{{molK}}\)

0.0330

0.030

0.0280

0.0249

0.026

0.036

0.0276

0.030

  1. The results of the proposed AEGCNN-MTL model (Ours) are compared against several representative baselines, including SchNet, MEGNet, PhysNet, DimeNet, ET, R-MAT, and GeoT. Reported MAEs are computed using consistent training/test splits of QM9, with bold numbers highlighting the best performance for each property.
Back to article page