Vibrational Frequencies of Molecular Systems using High Dimensional Neural Network Potentials

Abstract

Computational Chemistry is an important field in chemistry which looks for solu- tions for several questions such as reaction mechanisms, design of experiments and understanding fundamental properties of molecules. For molecular systems, usually quantum chemical methods are used. These methods give highly accurate results albeit with high computational costs. They can be used to calculate several proper- ties like energies, forces and also vibrational spectra. When it comes to vibrational computations, there are two limiting factors; the level of electronic structure meth- ods and the level of vibrational treatment. A highly accurate Potential Energy Surface (PES) is needed to compute high quality vibrational frequencies. Machine Learning Potentials (MLPs) which have become increasingly popular in chemistry and material science can help in bridging the gap between accuracy and cost. Here, High-Dimensional Neural Network Potentials (HDNNPs) are the MLPs of choice used to construct an accurate Potential Energy Surface for use in vibrational spec- troscopy. The endeavour is to construct an HDNNP fitted to a computationally expensive Cou- pled Cluster method starting from a small molecule which is Formic Acid Monomer and then increase the system size to the Formic Acid Dimer. The constructed PES will be used to calculate vibrational frequencies at harmonic and anharmonic levels and also benchmark them against experimental and theoretical vibrational frequen- cies. The HDNNP has further benefits of accurately representing the energies and dynamics of the system at hand at low cost. The work presents a methodology to construct a High-Dimensional Neural Network Potential for use in vibrational spectroscopy. The PES is constructed systemati- cally with proper analysis and validation steps to reach a predefined threshold of 10 cm−1 for harmonic frequencies. Additionally, the HDNNP for Formic Acid Dimer is validated by computing anharmonic frequencies. Concurrently, it tests the capa- bilities of a High-Dimensional Neural Network in representing the fine details of the potential.