Diksha Kadre

Identification of crystal polymorphs of natural as hydrates using machine learning

The ability of the water molecules to form complex hydrogen bonded structures results in the formation of cages called host, that can fit up the other small or large gas molecule, for example methane in our case, also known as guests within it. Such structures are crystalline in nature and are termed as the natural gas hydrates. With the advances in computational techniques and the increase in computational power, it is now possible simulate such systems with higher accuracy. However, it is difficult to understand the mechanism of nucleation using molecular dynamics (MD) simulations. Furthermore, it is also important to identify crystal polymorphs for controlling nucleation by understanding their thermodynamic stabilities and kinetic properties as it aids in designing strategies to manipulate nucleation and characterizing synthesized materials. Many algorithms were proposed that can identify ice [1] or hydrates, or both [2] but failed to differentiate among their polymorphs. We propose a supervised machine learning method to analyse the local structure around a central atom and classify different crystal polymorphs of water in presence of natural gas (methane), for example: natural gas hydrates, ice and water during the nucleation of gas hydrates. The molecular arrangements seem to be better understood by computing and analysing bond order parameters, which has been proven to be an efficient in distinguishing different structures. In this work, we identify crystal polymorphs of natural gas hydrates by first simulating different structures of ice, water and natural gas hydrates using molecular dynamics simulation and then employing the machine learning algorithm that uses different sets order parameters for given crystal polymorphs. Our studies found that neighbouring environments to central atom plays an important role in classifying these polymorphs. In particular, the classification accuracy based on the order parameters computed using the vectors from the molecule to all the other neighbouring atoms is more than the ones computed using the vectors from the molecule to other molecules. We have also shown that our model classifies these crystal polymorphs with a very high accuracy (test accuracy > 99%) with only few features.