Date of Award
8-2019
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Chemical and Biomolecular Engineering
Committee Member
Sapna Sarupria, Committee Chair
Committee Member
David A Bruce
Committee Member
Steven J Stuart
Committee Member
Rachel B Getman
Committee Member
Joseph K Scott
Abstract
Crystallization is a fundamental physical phenomenon with broad impacts in science and engineering. Nonetheless, mechanisms of crystallization in many systems remain incompletely understood. Molecular dynamics (MD) simulations are a powerful computational technique that, in principle, are well-suited to offer insights into the mechanisms of crystallization. Unfortunately, the waiting time required to observe crystal nucleation in simulated systems often falls far beyond the limits of modern MD simulations. This rare-event problem is the primary barrier to simulation studies of crystallization in complex systems. This dissertation takes a combined approach to advance simulation studies of nucleation in complex systems. First, we apply existing tools to a challenging problem — clathrate hydrate nucleation. We then use methods development, software development, and machine learning to address the specific challenges to simulation studies of crystallization posed by the rare-event problem.
Clathrate hydrate formation is an exemplar of crystallization in complex systems. Nucleation of clathrate hydrates generally occurs in systems with interfaces, and even homogeneous hydrate nucleation is inherently a multicomponent process. We address two aspects of clathrate hydrate nucleation which are not well-studied. The first aspect is the effects of interfaces on clathrate hydrate nucleation. Interfaces are common in hydrate systems, yet there are few studies probing the effects of interfaces on clathrate hydrate nucleation. We find that nucleation occurs through a homogeneous mechanism near model hydrophobic and hydrophilic surfaces. The only effect of the surfaces is through a partitioning of guest molecules which results in aggregation of guest molecules at the hydrophobic surface. The second aspect is the effect of guest solubility in water on the homogeneous nucleation mechanism. Experiments show that soluble guests act as strong promoter molecules for hydrate formation, but the molecular mechanisms of this effect are unclear. We apply forward flux sampling (FFS) and a committor analysis to identify good approximations of the reaction coordinate for homogeneous nucleation of hydrates formed from a water-soluble guest molecule. Our results suggest the possibility that the nucleation mechanism for hydrates formed from water-soluble guest molecules is different than the nucleation mechanism for hydrates formed from sparingly soluble guest molecules.
FFS studies of crystal nucleation can require hundreds of thousands of individual MD simulations. For complex systems, these simulations easily generate terabytes of intermediate data. Furthermore, each simulation must be completed, analyzed, and individually processed based upon the behavior of the system. The scale of these calculations thus quickly exceeds the practical limits of traditional scripting tools (e.g., bash). In order to apply FFS to study clathrate hydrate nucleation we developed a software package, SAFFIRE. SAFFIRE automates and manages FFS with a user-friendly interface. It is compatible with any simulation software and/or analysis codes. Since SAFFIRE is built on the Hadoop framework, it easily scales to tens or hundreds of nodes. SAFFIRE can be deployed on commodity computing clusters such as the Palmetto cluster at Clemson University or XSEDE resources.
Studying crystal nucleation in simulations generally requires selecting an order parameter for advanced sampling a priori. This is particularly challenging since one of the very goals of the study itself may be to elucidate the nucleation mechanism, and thus order parameters that provide a good description of the nucleation process. Furthermore, despite many strengths of FFS, it is somewhat more sensitive to the choice of order parameter than some other advanced sampling methods. To address these challenges, we develop a new method, contour forward flux sampling (cFFS), to perform FFS with multiple order parameters simultaneously. cFFS places nonlinear interfaces on-the-fly from the collective progress of the simulations, without any prior knowledge of the energy landscape or appropriate combination of order parameters. cFFS thus allows testing multiple prospective order parameters on-the-fly.
Order parameters clearly play a key role in simulation studies of crystal nucleation. However, developing new order parameters is difficult and time consuming. Using ideas from computer vision, we adapt a specific type of neural network called a PointNet to identify local structural environments (e.g., crystalline environments) in molecular simulations. Our approach requires no system-specific feature engineering and operates on the raw output of the simulations, i.e., atomic positions. We demonstrate the method on crystal structure identification in Lennard-Jones, water, and mesophase systems. The method can even predict the crystal phases of atoms near external interfaces. We demonstrate the versatility of our approach by using our method to identify surface hydrophobicity based solely upon positions and orientations of nearby water molecules. Our results suggest the approach will be broadly applicable to many types of local structure in simulations.
We address several interdependent challenges to studying crystallization in molecular simulations by combining software development, method development, and machine learning. While motivated by specific challenges identified during studies of clathrate hydrate nucleation, these contributions help extend the applicability of molecular simulations to crystal nucleation in a broad variety of systems. The next step of the development cycle is to apply these methods on complex systems to motivate further improvements. We believe that continued integration of software, methods, and machine learning will prove a fruitful framework for improving molecular simulations of crystal nucleation.
Recommended Citation
DeFever, Ryan Scott, "Advancing Molecular Simulations of Crystal Nucleation: Applications to Clathrate Hydrates" (2019). All Dissertations. 2470.
https://tigerprints.clemson.edu/all_dissertations/2470