Please use this identifier to cite or link to this item:
http://arks.princeton.edu/ark:/88435/dsp01nk322h596
Title: | First-Principles-Based Machine Learning Models for Fluid Phase Behavior |
Authors: | Nicola Barbosa Muniz, Maria Carolina |
Advisors: | PanagiotopoulosCar, AthanassiosRoberto |
Contributors: | Chemical and Biological Engineering Department |
Keywords: | Fluid Phase Behavior Machine Learning Potentials Molecular Simulations Statistical Mechanics |
Subjects: | Chemical engineering |
Issue Date: | 2023 |
Publisher: | Princeton, NJ : Princeton University |
Abstract: | Constructing accurate and efficient models of fluids such as water and CO2 for molecular simulations is of great interest to several industrial applications, given the limitations and challenges of experimental studies. Fluid phase behavior, specifically the vapor-liquid equilibrium (VLE), is involved in many energy and sustainability processes, from steam turbines for electricity generation to carbon capture and sequestration.As a way of developing fluid models that can be efficiently applied in large-scale molecular dynamics (MD) simulations, we turn to machine learning (ML) techniques to construct data-driven models that can reproduce potential energy surfaces based on ab initio methods or any reference method of choice. We start by first evaluating a known limitation of these short-ranged ML models: the localized representation and lack of explicit long-range interactions. We investigate the role of long-range electrostatics interactions in short-ranged ML models when predicting thermodynamic bulk liquid and vapor properties and VLE coexistence curves of water. After a better understanding of these effects, we switch to analyzing the performance of a promising many-body potential for water, constructed based on high level ab initio data, in reproducing VLE properties. Since MB-pol yields excellent agreement with experimental results for several properties and phases of water including VLE, but has an associated high computational cost, we construct a ML model based on this potential to expand its applicability in simulations. We also analyze how a ML model recently constructed based on the SCAN approximation of density functional theory (DFT) performs in the vapor-liquid interfacial region, which has not been used in its training. Lastly, we develop ML models for CO2, based on four different DFT exchange-correlation functionals. We assess how they perform in calculating bulk liquid densities, VLE properties and also transport properties of CO2. The studies presented in this dissertation shed light on to the capabilities and limitations of ML models, as well as their reference methods, in predicting thermodynamics and phase behavior of fluids. Furthermore, these models can provide an avenue for studying fluid phenomena from a first principles-based perspective. |
URI: | http://arks.princeton.edu/ark:/88435/dsp01nk322h596 |
Type of Material: | Academic dissertations (Ph.D.) |
Language: | en |
Appears in Collections: | Chemical and Biological Engineering |
Files in This Item:
File | Description | Size | Format | |
---|---|---|---|---|
NicolaBarbosaMuniz_princeton_0181D_14541.pdf | 4.84 MB | Adobe PDF | View/Download |
Items in Dataspace are protected by copyright, with all rights reserved, unless otherwise indicated.