Modelování molekulových interakcí pomocí neuronových sítí

Abstract

The forces acting between the particles during their interactions can be described by potential energy surfaces. The aim of this work is to evaluate whether it is possible to represent these surfaces using neural networks with sufficient precision.

For verification, I used two analytical models, namely Lennard-Jones potential and Morse potential. Surfaces were represented by feed-forward neural networks that were trained using back-propagation method.

In this work, such neural networks have been created that can predict potential energy from a relatively small set of input data. The resulting neural networks are also able of an approximate extrapolation of potential surfaces in large distances between particles outside the range of training data.

The results obtained in this thesis will serve as the basis for further research, which will deal with representations of larger systems. Thanks to this research, it will be possible to efficiently represent surfaces and predict potential energy values even in domains where ab initio methods have convergence problems or where their are computationally too demanding.