Inverse Infrared Spectroscopy With Bayesian Optimization
Jezrielle Annis, Texas A&M University
Infrared spectroscopy offers a direct, atomistic level probe of molecular environments. While high-level theoretical calculations are often required to correctly interpret and assign these spectra, the proper assignment of common CH and OH stretching modes requires the inclusion of anharmonic effects. For large systems these effects become infeasible to calculate due to the scale of calculations.
Here, we employ machine learning methods, specifically Bayesian-optimization to automatically find the anharmonic effects in highly anharmonic spectra. We benchmark against both high-level theoretical calculations and experimental spectra from the literature. This approach requires no prior training data, and only requires a single low-cost harmonic frequency calculation and a physically motivated, a priori form of the Hamiltonian to initiate. This “Inverse Spectroscopy” approach actively learns the anharmonic Hamiltonian from the target spectra with minimal computational cost. Here, we demonstrate the approach on several problem types, including benchmark systems and experimental spectra from literature.
Abstract Author(s): Jezrielle R. Mildon, Daniel P. Tabor