Quantum mechanical and machine learning prediction of rotational energy barriers in halogenated aromatic alcohols

TL;DR

This paper uses quantum mechanics and machine learning to predict how substituents affect the energy barriers for rotation in aromatic alcohols.

Contribution

A machine learning model outperforms chemical intuition in predicting rotational energy barriers in halogenated aromatic alcohols.

Findings

Machine learning models trained on DFT data predict rotational energy barriers better than chemical intuition-based correlations.

Models trained separately on pyrenols, anthranols, or phenols perform better than those trained on all compound classes together.

Models trained on single-substituent compounds generalize well to mixed-substituent compounds.

Abstract

Rotation about a chemical bond is important in many chemical processes and can be influenced by neighboring substituents on a molecule. Rotational energy barriers can be predicted by density functional theory (DFT) calculations. Here, we specifically explore how substituents influence the barrier to rotation about the C-O bond in symmetrically halogenated aromatic alcohols. A machine learning model was trained on the DFT-calculated rotational energies and was found to do a good job predicting rotational energy barriers from the electronegativity, atomic radius, and Hammett constant for each substituent. The machine learning model was found to perform better when it was trained separately on pyrenols, anthranols, or phenols than when it was trained on all classes of compounds together. Even though the models were trained on compounds containing only one kind of substituent, they were found to perform similarly well on compounds containing mixed substituents. Machine learning was able to predict the rotational energy barrier heights better than correlations among parameters that would be expected to be relevant based on chemical intuition.

DFT calculations were done with Gaussian 16 software at the B3LYP/6–311 + G(d.p) level of theory. Machine learning was done using the classification and regression training (caret) package in R version 4.4.0.

The online version contains supplementary material available at 10.1007/s00894-025-06321-y.

Figures