Figure 29.

Performance of ML models for atomization energies of small organic molecules and radicals. (a) Learning curves for the QM9 dataset,²³² using a variety of representations and regression methods. For each value of the training set size, we show the mean absolute error (MAE) evaluated on the test set which consists of the remaining structures from the full dataset. Models based on FCHL (2018),²³³ SOAP (2018),⁶⁹ aSLATM,²³⁴ Coulomb Matrix (CM),²³⁵ and Bag-of-bonds (BOB)²³⁶ representations use Gaussian process/kernel ridge regression, whereas NICE⁷³ and MTP⁶⁷ use linear ridge regression, and SchNet²³⁷ and PhysNet²³⁸ are graph neural networks. GM-sNN uses a representation similar in spirit to MTP but based on a Gaussian radial basis set and a feed-forward neural network for regression.²³⁹ (b) Learning curve for the Rad-6 dataset.²⁴⁰ Example species are shown including a radical species, which actually account for over 90% of the total dataset (reprinted from ref (240); original work published under the CC BY 4.0 license; https://creativecommons.org/licenses/by/4.0/).