Discussion

Effects of Conformer Energy Ranges on Accuracy Metrics

Previous work has suggested that the poor correlations found between force field and semiempirical methods are derived from the small number of low-energy conformers considered in this benchmark.\cite{Sharapa_2018}  Certainly, one might imagine that when considering multiple geometries with only small differences in energies, random errors are magnified. Figure \ref{361711} illustrates a histogram of the ranges in DLPNO-CCSD(T) energies across the molecules considered. Despite the small ranges in energies, there is little correlation between the energy range of a molecule and the accuracy metrics of a particular method. This suggests no bias from the small energy windows used in this benchmark set.
Figure \ref{584948}  indicates there is no correlation between R2 and the energy window of the conformers. The ML methods have a relatively even distribution of R2 across the energy window indicating that random errors in the model may have more of an impact on performance than the size of the energy window.