The X1 family of methods that combines B3LYP with neural network corrections for an accurate yet efficient prediction of thermochemistry

Abstract

B3LYP is currently the most widely used density functional approximation, while the X1 family of methods, namely X1, X1s, and X1se, is a set of neural network-based methods that systematically correct the B3LYP errors. The performance of the X1 family of methods in the prediction of heats of formation (HOFs), bond dissociation enthalpies (BDEs), heats of isomerization (HOIs), and so forth, is summarized against some well-established benchmarking datasets. X1 significantly eliminates the notorious size-dependent errors of B3LYP in prediction of HOFs for larger molecules. X1s further exhibits a significant improvement for BDE calculations. X1se continues to improve its predecessors on HOIs. Such a progressive improvement relies on the increasingly comprehensive descriptors. Limitations of the present approaches and the direction for future improvements are discussed.

Accurate heats of formation of polycyclic saturated hydrocarbons predicted by using the XYG3 type of doubly hybrid functionals

“…Some effective correction schemes have been developed to improve the accuracy of B3LYP HOFs . Previously, we have compared the HOFs via the atomization procedure from B3LYP‐D2, B3LYP‐D3, and B3LYP‐D3BJ using the basis set of 6‐311 + G(3df,2p).…”

“…Machine learning is another way to improve the accuracy of B3LYP HOFs . For example, the X1 and X1s methods applied some neural‐network corrections to the B3LYP HOFs, where MADs of 6.7 and 5.9 kcal/mol, respectively, have been achieved for this PSH set.…”

“…Furthermore, it has been found that the presence of some particular structural and stereoelectronic features in an organic compound can lead to substantially increased errors for an approximate DFT method, which prevents it from being used as a black‐box method in theoretical organic chemistry to explore the important stereoelectronic effects . Many correction schemes have been developed to overcome such shortcomings while new density functionals of higher accuracy have also been proposed to go beyond B3LYP.…”

Improving B3LYP heats of formation with three‐dimensional molecular descriptors

“…Refs. [ ]). The DFA errors were related to certain characteristic properties (or descriptors), leading to methods based on parametrization of atomic energies, bond/group additivity corrections, and corrections considering spin multiplicities and charges, etc.…”

Special issue on machine learning and quantum mechanics

“…In their contribution, Albert P. Bartók and Gábor Csányi take the reader on a tour through their Gaussian approximation potentials approach for potential energy surface interpolation, and Paul L.A. Popelier summarizes progress on the development of a GPR potential for peptides and proteins based on Quantum Chemical Topology . An example of thermochemical property predictions across different molecules can be found in the article by Jianming Wu, Yuwei Zhou, and Xin Xu, who use ANNs to statistically correct DFT/B3LYP predictions with respect to experimental values …”