Files
Abstract
Accurate theoretical prediction of molecular vibrational frequencies is important for informing chemical experiments. Coupled-cluster theory with a second-order perturbative treatment of anharmonicity (VPT2) can yield accurate frequencies to within 10-20 cm−1, at least for semi-rigid molecules without resonances. Outside of these cases, however, VPT2 can be rendered useless for some of the most challenging chemical systems. Additionally, spectroscopy accuracy (within 1 cm−1) remains out of reach for the vast majority of vibrational spectra. Herein, I investigate the methylene amidogen radical (H2CN) as a case study for current theoretical techniques that treat vibrational motion. I analyze the error sensitivity of our state-of-the-art theoretical treatment and find that its limits of accuracy lie in the VPT2 treatment of anharmonicity. A next step toward spectroscopic accuracy might be vibrational configuration interaction, but this relies on the availability of an accurate energy surface, which is often prohibitively expensive. Motivated by this need, I discuss a machine learning technique called autoregressive Gaussian process modeling, which can reduce the computational cost of obtaining training data by leveraging relationships between low- and high-accuracy models. I apply this approach to the prediction of a chemical energy surface for the first time. The initial benchmarks presented here suggest that it can significantly improve learning efficiency.