Table 2 Single-task versus multi-task embeddings on mean absolute error (MAE) for formation energy (eV/atom) and bandgap (eV) and R2

From: Transformer-generated atomic embeddings to enhance prediction accuracy of crystal properties with machine learning

Target

None-CG

\({{{\rm{CT}}}}^{{E}_{f}}\)-CG

\({{{\rm{CT}}}}^{{E}_{g}}\)-CG

CTMT@2p-CG

CTMT@3p-CG

CTMT@4p-CG

MAE(Ef)

0.083

0.071

0.078

0.068

0.069

0.068

R2(Ef)

0.984

0.987

0.983

0.987

0.987

0.986

MAE(Eg)

0.384

0.383

0.359

0.357

0.356

0.367

R2(Eg)

0.845

0.845

0.850

0.849

0.851

0.847

  1. CG indicates CGCNN. None means no embeddings are used, Ef and Eg denotes embeddings trained on the corresponding target. MT@np demotes embeddings trained with multi-task learning on n properties.
Back to article page