Table 3 Comparison of test-set MAE values for the three winning models trained for a re-partitioned dataset containing five of six lattice symmetries (2384 samples) and a test set (616 samples) comprising only one lattice symmetry (\({\mathrm{Ia}}\bar 3\))

From: Crowd-sourcing materials-science challenges with the NOMAD 2018 Kaggle competition

Representation

Regressor

Formation energy (meV/cation)

Bandgap energy (meV)

4-gram

KRR

53

179

c/BOP

LGBM

40a/36b

180a/111b

SOAP

NN

132

527

  1. aFeature selection and model hyperoptimization according to fivefold CV with splits generated randomly
  2. bFeature selection and model hyperoptimization according to fivefold CV with splits generated based on the spacegroup number