Table 2 Mean absolute errors for total energies (\(U_0\) and E in kcal/mol), HOMO and LUMO energies (in eV) using different ML methods on different training and prediction datasets
From: Dataset’s chemical diversity limits the generalizability of machine learning predictions
Method | Train | Test | \(U_0\) | E | HOMO | LUMO |
|---|---|---|---|---|---|---|
EN (CM) | QM9 | QM9 | 21.1 | 22.0 | 0.34 | 0.64 |
KRR (CM) | QM9 | QM9 | 4.9 | 5.2 | 0.18 | 0.25 |
SchNet | QM9 | QM9 | 0.3 | 1.0 | 0.04 | 0.03 |
EN (CM) | PC9 | PC9 | – | 38.2 | 0.47 | 0.66 |
KRR (CM) | PC9 | PC9 | – | 22.8 | 0.31 | 0.36 |
SchNet | PC9 | PC9 | – | 1.6 | 0.06 | 0.05 |
SchNet | QM9 | PC9(A) | – | 3.0 | 0.07 | 0.06 |
SchNet | QM9 | PC9(B) | – | 8.9 | 0.33 | 0.27 |
SchNet | PC9 | QM9(A) | – | 3.4 | 0.05 | 0.05 |
SchNet | PC9 | QM9(B) | – | 4.2 | 0.12 | 0.11 |