Fig. 6
Atomization energy learning curves on the subset of 258k unique constitutional isomers from VQM24 and the QM9 dataset. Solid lines indicate ML models employing Kernel Ridge Regression (KRR) while dashed lines indicate Graph Neural Networks (GNN). Representations used alongside KRR models are : Coulomb matrix1 (CM), atom-centered symmetry functions97 (ACSF), many-body tensor representation98 (LMBTR), Faber-Christensen-Huang-Lilienfeld 1999,100 (FCHL19), convolutional many-body distribution functionals94,95 (cMBDF/MBDF). GNNs employ the equivariant PaiNN90 and invariant SchNet89 architectures. Test set size in both cases was 10,000 randomly-selected molecules. Plots show average of 5 such runs.
