Machine Learning for Exploration of Silica Structures Using Molecular Dynamics

Abstract

We use molecular dynamics and machine learning for exploring structures in alpha quartz, a crystalline form of silica. This material is simulated at the molecular level, and structures are created by carving out atoms from different locations in the material. By removing atoms, we can change the physical properties of the material, such as the yield stress. We use simplex noise (a type of pseudo-random noise) to sample different structures in alpha quartz, and we demonstrate that these structures represent materials with a wide range of yield stresses. The yield stresses are obtained by simulating tensile strain on the systems. Using data from many systems of alpha quartz, we train a convolutional neural network (CNN) for predicting the yield stress of alpha quartz structures. This CNN is used in a search algorithm that tries to find structures that are stronger than the structures used in training the CNN. The structures are restricted so that they have roughly the same porosity. The search algorithm was successful in finding stronger structures, and the yield stress of the strongest material is estimated at 3.412 GPa. We show that the search algorithm can be modified to instead look for weaker materials, and also that it can be used to search for materials with a specific yield stress. The weakest material found was estimated to have a yield stress of 2.103 GPa. We find that the strongest alpha quartz structures follow a distinct pattern. The cuts in these structures are arranged in a hexagonal shape, thus maximizing the minimum distance between individual cuts. In weak structures, we find that there are usually many cuts that are placed close to each other. These results indicate that the yield stress of alpha quartz has a notable dependence on how the cuts are placed relative to each other. To maximize the yield stress, cuts should be placed as far apart as possible.