Table 2 Test set results for pooling methods and HOMO energy prediction on the OE62 dataset (SchNet).

From: Modelling local and general quantum mechanical properties with attention-based pooling

Readout

MAE

RMSE

# ABP parameters

Sum

0.2656 ± 0.0177

0.4032 ± 0.0168

N/A

Average

0.1437 ± 0.0016

0.2043 ± 0.0009

N/A

OWA

0.1135 ± 0.0019

0.1670 ± 0.0021

N/A

ABP(64, 4, 2)

0.1158 ± 0.0020

0.1697 ± 0.0021

995,456

ABP(64, 8, 2)

0.1130 ± 0.0019

0.1660 ± 0.0028

3,694,720

ABP(64, 16, 2)

0.1119 ± 0.0022

0.1655 ± 0.0045

14,205,056

ABP(64, 16, 3)

0.1124 ± 0.0006

0.1648 ± 0.0015

18,403,456

  1. Test MAE and RMSE (mean ± standard deviation from 5 data random splits) for SchNet-based HOMO energy prediction on the OE62 dataset, including the number of learnable parameters for the attention-based pooling (ABP). The ABP configuration is reported as ‘ABP(embedding size, number of attention heads, number of SABs)’. The number of learnable parameters for the underlying SchNet model (not including the readout) is 480,002. The smallest MAE/RMSE values and table headers are highlighted in bold. The unit used for energy is eV, as used by Chen et al.27.
Back to article page