Polarizability Models for Simulating Raman Spectra with Molecular Dynamics

The progress of technology constantly requires the development of new materials used for solar cells or batteries. Experimental discoveries can be guided by computer simulations, although they can prove computationally expensive and sometimes have limited accuracy. For example, simulating Raman spectra is extremely demanding, while it is a commonly used experimental technique. Due to their fast development in recent years, machine learning algorithms can now be used to speed up simulations and reduce the computational load.

This doctoral thesis explores how machine learning methods could be used to improve the simulation of Raman spectra. These are obtained by combining state-of-the-art machine learning force fields with various polarizability models, ranging from simple empirical models to more complex neural networks. Although all considered models lead to satisfactory prediction of the Raman spectrum, empirical models are found to be less accurate yet faster than machine learning polarizability models.

Models are then applied to various scientifically relevant systems, including solid state semiconductors, 2-dimensional materials and amino acid chains. By making use of the high efficiency of machine learning force fields and polarizability models developed in the thesis, large-scale simulations are performed, allowing to obtain new insights in the Raman spectra of each material.