Machine-learning techniques for geometry optimization

Abstract

Geometry optimization in computational chemistry is still a challenging task. The bottleneck is the computationally expensive ab initio calculations. Thus reducing the total amount of these calculations to accelerate minimization and transition state search is essential. In recent years machine-learning techniques, like Gaussian process regression or neural networks, became popular among scientists. These can be used to calculate the surrogate surface of the potential energy surface and perform geometry optimization there. Another important aspect of geometry optimization is the choice of coordinate system. While Cartesian coordinates describe uniquely a molecule, they are highly coupled. To reduce the coupling between the coordinates, the so-called internal coordinates were developed a long time ago. In addition, these coordinates are non-redundant. With these types of coordinates, a speedup of geometry optimization was obtained. Combining internal coordinates with machine-learning technique has thus the potential to significantly improve geometry optimization. In this thesis, the improvement is demonstrated.