Table 7 Energy errors in meV on a reference dataset used to train a CE model for Ag-Pd.

From: Machine-learned interatomic potentials for alloys and alloy phase diagrams

 

GAP

MTP

CE

Training error (meV)

2.7

1.9

1.6

Validation error (meV)

5.7

4.2

2.1

Basis size

1075

569

41

  1. CE cluster expansion.
  2. Details on building the dataset and model are given in ref. 81. While the CE model is slightly better, both GAP and MTP perform well in that region of configuration space where the CE model is valid. We also include the number of basis functions used by each models as a surrogate for performance. Because the CE basis is well-suited to on-lattice modeling, comparatively few basis functions are needed for its predictions. Note that the reported training error is with respect to the dataset used to train the model; since CE does not train on off-lattice data, its training dataset is different from GAP and MTP.
Back to article page