Abstract

The assignment of spectral lines in both the visible region and the UV region has important applications in the study of astral objects such as stars and exoplanets, or, closer to us, in the study of the molecular composition of the terrestrial atmosphere. However, the correct assignment of spectral lines cannot solely rely on experimental results, since those are not always available, making theoretical predictions necessary.

To this end, we extended the use of machine learning techniques, more specifically of Gaussian regression processes, in PES prediction to the use of composite Gaussian regression processes (c-GP) trained at different levels of theory, with different training sets, as well as different coordinates systems. The PES is then given as a sum of probabilistic prediction corresponding to dense training sets at low levels of theory and sparser training sets on computationally expensive deterministic (or stochastic) methods.

The study of the most abundant cation in the Universe (and simplest polyatomic cation), H3+, is used to assess the performance of c-GPs.