Compare ML methods to DFT and DLPNO methods. (ANI-1x does not include dispersion correction.. may be one reason ANI-1ccx is better.)
Table \ref{310223} and Figure \ref{444292} show the ANI symmetry function ML methods, ANI-1x, ANI-1ccx, and ANI-2x, performing similarly to tight binding semiempirical methods in both accuracy and speed. ANI-1cxx outperforms both the ANI-1x and ANI-2x models that do not contain dispersion corrections. The inclusion of dispersion correction for DFT methods was shown to be beneficial as they outperformed their non-dispersion corrected counterparts as seen in Table [X]. It is conceivable that an ANI-2ccx or ANI-3 model with dispersion correction may provide additional improvements to the current ANI-2x model. The performance of the bag-of-features models, while faster than the ANI symmetry function models, were more comparable to the accuracy of force field methods. The inclusion of additional information to the descriptor such as three and four-body interactions and atom typing were beneficial to the bag-of-features models, the accuracy pales in comparison to the ANI symmetry function models.
[ Despite training on DFT and coupled-cluster energies, ML methods still not as accurate for conformer energies.. why? Maybe some random errors, maybe don't include multiple minima in parameterization, so there's a bias in training? ]
Despite training on DFT and coupled-cluster energies, ML methods still are not as accurate for predicting conformer energies. Figure (Ewindow figure) indicates there is no correlation between R2 and the energy window of the conformer and the distribution of the 
[E window Figure showing random error for ML method(s)?]