Structural Characterization of Water in Different Thermodynamic Conditions Through Unsupervised Learning

The description of water’s microscopic structure and the relation to its thermody- namic and dynamic anomalies remain an open challenge. In this thesis we approach this challenge in an agnostic manner. We deploy a variety of unsupervised learning algorithms; starting by encoding the water environments into symmetry invariant atomistic descriptors, computing their intrinsic manifold, and subsequently comput- ing the free energy landscape to uncover the relevant microscopic structures of water near the Liquid-Liquid Critical (LLCP). The microscopic signatures underlying the LLCP has been rationalized in terms of two locally stable competing structures char- acteristic of the High Density (HD) and the Low Density (LD) liquid phases. In contrast to this notion, we agnostically discover non-local domains that are relevant to characterize the LD and HD phases. These domains span length scales of up to 1 nm. We also investigate the coupling between different Structural Parameters used in the study of water in these conditions and the macroscopic density. Confirming our interpretations, the descriptors are maximally coupled to the density only on length scales of 1 nm and beyond. Additionally, we applied our unsupervised protocol to automatically discover the microscopic fingerprints that characterize the fluctuations of water near extended hy- drophobic interfaces. We show that the surface structure of water near these interfaces is quite heterogeneous and orientationally anisotropic only up to 3.5 ˚A away from the intrinsic surface; it is a mixture of both bulk water environments as well as interfacial water environments whose hydrogen bond patterns are very different from the bulk. The methods employed here details a protocol that can be used to study the fluc- tuations of water in other conditions such as in confinement or at the interface with other biologically relevant macro-molecules like proteins.