Ethereal AI: Infrared Spectra of Polycyclic Aromatic Hydrocarbons with Machine Learning DFT Scaling Factors

Polycyclic aromatic hydrocarbons (PAHs) are of interest
in astrochemistry
and are studied through infrared (IR) signals from space. Deconvoluting
the observed IR of PAHs often relies on computational chemistry for
references. Current practice typically uses IR predicted by density
functional theory adjusted with scaling factors to account for anharmonicity
and basis set limitations. We present a machine learning approach
to scaling factors that adjusts each predicted IR frequency individually
based on computed frequencies, intensities, reduced masses, and force
constants. Our machine learning model is able to obtain a mean absolute
error (MAE) and maximum error of 5 and 13 cm–1,
respectively, a drastic improvement over a current-practice scaling
factor approach with an MAE and maximum error of 10 and 23 cm–1, respectively. Using machine learning to determine
scaling of DFT-generated IR spectra shows promise for enabling higher
quality predictions on larger systems than those that have previously
been possible.