Table 2 Elpasolite formation energy prediction results after ten-fold cross-validation; mean best formation energy MAE on the test set after 200 epochs of training in each fold.

From: Distributed representations of atoms and materials for machine learning

Representation

Dim

MAE (eV/atom)

Atom2Vec

30

0.1477 ± 0.0078

SkipAtom

30

0.1183 ± 0.0050

Random

30

0.1701 ± 0.0081

Atom2Vec

86

0.1242 ± 0.0066

One-hot

86

0.1218 ± 0.0085

SkipAtom

86

0.1126 ± 0.0078

Random

86

0.1190 ± 0.0085

Mat2Vec

200

0.1126 ± 0.0058

SkipAtom

200

0.1089 ± 0.0061

Random

200

0.1158 ± 0.0050

  1. Bold value represents the best result.
  2. Batch size was 32, learning rate was 0.001. Note that Dim refers to the dimensionality of the atom vector; the size of the input vector is 4 × Dim. All results were generated using the same procedure on identical train/test folds.
Back to article page