Infrared spectra of neutral polycyclic aromatic hydrocarbons by machine learning

TL;DR

This paper introduces a machine learning approach using neural networks to accurately predict infrared spectra of polycyclic aromatic hydrocarbons, achieving transferability from small to large molecules with reduced computational cost.

Contribution

The study develops a neural network-based potential energy surface and dipole mapping that accurately predicts IR spectra of large PAHs from training on small molecules, demonstrating transferability.

Findings

Neural networks accurately reproduce IR spectra of small PAHs.

The approach successfully predicts spectra of larger PAHs not in training.

Transferability reduces computational costs compared to traditional methods.

Abstract

The Interest in polycyclic aromatic hydrocarbons (PAHs) spans numerous fields and infrared spectroscopy is usually the method of choice to disentangle their molecular structure. In order to compute vibrational frequencies, numerous theoretical studies employ either quantum calculation methods, or empirical potentials, but it remains difficult to combine the accuracy of the first approach with the computational cost of the second. In this work, we employed Machine Learning techniques to develop a potential energy surface and a dipole mapping based on an artificial neural network (ANN) architecture. Altogether, while trained on only 11 small PAH molecules, the obtained ANNs are able to retrieve the infrared spectra of those small molecules, but more importantly of 8 large PAHs different from the training set, thus demonstrating the transferability of our approach.