Efficient Composite\nInfrared Spectroscopy: Combining\nthe Double-Harmonic Approximation with Machine Learning Potentials

Vibrational spectroscopy\nis a cornerstone technique for\nmolecular\ncharacterization and offers an ideal target for the computational\ninvestigation of molecular materials. Building on previous comprehensive\nassessments of efficient methods for infrared (IR) spectroscopy, this\nstudy investigates the predictive accuracy and computational efficiency\nof gas-phase IR spectra calculations, accessible through a combination\nof modern semiempirical quantum mechanical and transferable machine\nlearning potentials. A composite approach for IR spectra prediction\nbased on the double-harmonic approximation, utilizing harmonic vibrational\nfrequencies in combination squared derivatives of the molecular dipole\nmoment, is employed. This approach allows for methodical flexibility\nin the calculation of IR intensities from molecular dipoles and the\ncorresponding vibrational modes. Various methods are systematically\ntested to suggest a suitable protocol with an emphasis on computational\nefficiency. Among these methods, semiempirical extended tight-binding\n(xTB) models, classical charge equilibrium models, and machine learning\npotentials trained for dipole moment prediction are assessed across\na diverse data set of organic molecules. We particularly focus on\nthe recently reported foundational machine learning potential MACE-OFF23\nto address the accuracy limitations of conventional low-cost quantum\nmechanical and force-field methods. This study aims to establish a\nstandard for the efficient computational prediction of IR spectra,\nfacilitating the rapid and reliable identification of unknown compounds\nand advancing automated high-throughput analytical workflows in chemistry.