Table 3 Various embedding approaches comparison 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

CT-CGCNN

CTchem+coords-CGCNN

CTfreeze-CGCNN

MAE(Ef)

0.071

0.085

0.073

R2(Ef)

0.987

0.983

0.986

MAE(Eg)

0.359

0.395

0.358

R2(Ef)

0.850

0.834

0.851

  1. CT denotes embeddings trained on corresponding properties. CTchem+coords denotes atom and coordinates embeddings, while CTfreeze denotes embeddings with zero grad when training the back-end model.
Back to article page