ATOMISTIC LEVEL INVESTIGATION OF DISSOLUTION, PRECIPITATION AND SURFACE DIFFUSION DYNAMICS AT THE AQUEOUS INTERFACES USING REAXFF AND APPLICATION OF DEEP LEARNING TO FORCE FIELD PARAMETERIZATION
Open Access
- Author:
- Sengul, Mert Yigit
- Graduate Program:
- Materials Science and Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- February 18, 2020
- Committee Members:
- Clive A Randall, Dissertation Advisor/Co-Advisor
Clive A. Randall, Committee Chair/Co-Chair
Adrianus C Van Duin, Committee Chair/Co-Chair
Ismaila Dabo, Committee Member
Enrique Daniel Gomez, Outside Member
Adrianus C Van Duin, Dissertation Advisor/Co-Advisor
John C Mauro, Program Head/Chair - Keywords:
- Cold Sintering Process
ReaxFF
Aqueous Interfaces
Deep Learning
Machine Learning
Molecular Dynamics
Surface Diffusion
Langmuir Model - Abstract:
- Cold Sintering Process (CSP) densifies ceramics at extremely low temperatures compared to conventional sintering processes. A solvent is mixed with ceramics powder during the application of CSP, which drives the densification through dissolution and recrystallization. Several ceramics and composite systems have been successfully densified under cold sintering, yet the mechanism behind the extraordinary densification and grain growth observed under the cold sintering process of ceramics is unknown. In this study, we adapted ReaxFF interatomic potentials to study two metal-oxide/solvent interfaces, which are zinc oxide/acetic acid/water and Li1.3Al0.3Ti1.7(PO4)3 (LATP)-water. We investigated surface diffusion, adsorption, desorption and dissolution dynamics at these interfaces at the atomistic level. The ReaxFF carbon-oxygen-hydrogen force field was reoptimized to reproduce the experimental pKa value by simulations and was used to simulate acetic acid-water mixtures. The pKa value was calculated using a metadynamics method, which is an accelerated molecular dynamics (MD) procedure. The microscopic structures of acetic acid-water mixtures at different acidities were examined at ambient and supercritical conditions. When acetic acid is dominant in the mixture, cyclic dimers and chain structures between acetic acid molecules are observed, these dissociate at increased water content. The acid molecules released from these dimer and chain structures tend to stay in dipole-dipole interactions and this finding is in agreement with the experimental results. Our simulation results suggest that ReaxFF is an appropriate method for studying supercritical water/organic acid mixtures. Following the findings related to acetic acid-water mixtures, the zinc-oxygen-carbon-hydrogen force field was adapted to simulate zinc oxide-acetic acid-water interfaces. The atomistic level interactions were studied over a broad temperature range. According to our simulations, two different acetic acid dissociation mechanisms were observed, which are: 1) deprotonation to surface cation, which produces a terminal hydroxyl and (2) deprotonation to a bridging hydroxyl. In addition, acetic acid molecules decompose into carbon dioxide and formaldehyde at elevated temperatures, which is in agreement with experimental findings. The adsorption trends of water and acetic acid molecules observed in our simulations are consistent with a phenomenological Langmuir model. Recrystallization and surface diffusion of zinc cations at various acidic conditions at the ZnO-acetic acid-water interface were investigated by combining ReaxFF with experiments. According to our simulations, the presence of water in the system greatly enhances the surface/grain boundary diffusion during the evaporation of the solvent. In addition, our simulations revealed that the surface chemistry could have a significant effect on the activation of this accelerated surface diffusion mechanism; therefore, application of CSP requires careful selection of solvent. The ReaxFF potential that was developed to model LATP/water interfaces was used to investigate the aqueous dissolution dynamics in CSP. The simulations conducted in this work revealed that the dissolution at this interface is a sequentially dynamic process due to the different dissolution rates of the Li/Ti and Al cations. According to our simulations, the relatively early dissolution of lithium ions changes the charge distribution in the crystal phase, which creates a bonding transition between PO4 and AlO6. This transition results in the aluminum segregation and titanium-rich secondary phase formation. Alongside these MD studies, a deep learning-based high dimensional optimization procedure was developed to be used in ReaxFF force field parametrization, which is required to adapt ReaxFF potential to new materials systems in a more efficient way. The developed procedure starts with a Latin Hypercube Design (LHD) algorithm that is used to explore the parameter landscape extensively. The LHD passes the information about explored regions to the deep learning (DL) method for model training. The DL method is composed of a regression and classification model. The regression model finds the minimum low-discrepancy regions in parameter landscape, while the classification algorithm eliminates unfeasible regions, which originate from the unphysical atomistic interactions. As a result, the DL procedure constructs a more comprehensive understanding of a physically meaningful parameter landscape. The DL procedure was used to optimize a nickel-chromium ReaxFF force field, and the optimization quality was compared with the conventional gradient based method. According to this comparison, the DL procedure is capable of finding several better optimized force fields in a shorter period of time.