Abstract

Infrared spectroscopy is one of the most useful experimental techniques to identify and study molecular systems. However, the analysis and interpretation of those spectra can be very challenging for complex systems. In these cases theoretical spectroscopy based on molecular dynamics (MD) simulations can help unravel the microscopic details of the observed spectra. These simulations however require an accurate description of the electronic structure of the system to give reliable results. For many systems, the highest accuracy is currently provided by Coupled Cluster (CC) theory, sometimes refer to as the current “gold standard” of theoretical chemistry. However, the high computational cost of CC methods prohibit their use on-the-fly in MD simulation. Fortunately, Machine Learning (ML) methods have been actively developed in the last decade with great success to represent in particular the potential energy surface, as well as other properties, of chemical systems. In this talk we will show how one of these ML approaches, namely high-dimensional neural networks, can be used in order to describe both the PES and the electric dipole moments of molecular systems in order to run MD simulations essentially at CC level and obtain highly accurate IR spectra.