Machine learning reveals new insights into crystal nucleation in a Lennard-Jones fluid

Abstract

In this thesis we investigate the local crystalline structures on a single-particle level for Lennard-Jones particles during the process of crystal nucleation. We simulate crystal nucleation using Monte Carlo simulations in the NVT ensemble. During the simulations, the local structures around particles are analysed with a Principal Component Analysis and a neural network based classification algorithm. Both analyses show formation of a primarily face-centered cubic and hexagonal close-packed ordered crystal. The Principal Component Analysis suggests that crystal nucleation does not happen via body-centered cubic ordering, although we were unable to properly quantify the degree of body-centered cubic ordering. Lastly we attempt to improve the spatial resolution of the detection of local crystalline structures by using a neural network based autoencoder.