Uncertainty quantification in multiscale stochastic models of catalytic reactions
Date
2018
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Multiscale modeling, a key tool in probing the fundamentals of catalytic reactions,
has seen increased usage enabled by advances in computational hardware. Within
the multiscale modeling paradigm, kinetic Monte Carlo (KMC) is employed to simulate
chemical reaction networks, as mean-eld models often fail to provide a meaningful description
of the complex phenomena involved. Due to KMC's high computational cost
and stochastic noise, quantifying uncertainty for the purposes of rening the model
and assessing predictive reliability is dicult. Uncertainty arises from errors in input
parameters (parametric uncertainty) and assumptions made about the physical system
(model form uncertainty). ☐ In this thesis, we develop tools to quantify errors from each of the aforementioned
sources and make recommendations for model renement. We address parametric
uncertainty by developing ecient sensitivity analysis techniques, which identify the
most influential parameters. Likelihood ratio sensitivity analysis (LRSA) computes
all sensitivities without the need for additional runs, as required by nite dierence
methods, but encounters tremendous variance in systems with disparate time scales. To
overcome this limitation of LRSA, we derive mathematical theory that enables its use in
well-mixed multiscale KMC and implement the method in original software. The new
multiscale technique accurately computes sensitivities in a model system for which the
traditional LRSA performs poorly. To address spatial KMC, we develop acceleration
techniques and statistical criteria that ensure sucient sampling for LRSA. As a result,
LRSA can be applied to real chemistry. We apply our methodology to the water-gas
shift reaction on Pt, an important component of hydrogen production from biomass.
We address model form uncertainty by revisiting two common assumptions: the
structure of the catalyst surface is uniform and the identity of the active site is known. ☐ A framework for optimizing catalyst structure based on local descriptors is developed,
allowing for atom-by-atom design of defected surfaces and consequent improvements
in activity. In order to restrict our search to physically relevant structures, surface
energy is also computed. Activity is maximized and surface energy is minimized
simultaneously using multi-objective simulated annealing. A set of Pareto optimal
structures is found, oering targets for synthesis. We apply our approach to oxygen
reduction on Pt, the key reaction in automotive fuel cells. Our approach resolves
discrepancy between experiment and theory regarding the extent to which defects can
improve activity. We extend the approach to chemistries involving coupled active
sites, for which KMC simulation is needed. KMC simulation data from many dierent
structures is used to train a neural network for use as a surrogate model in the
optimization. The neural network is updated as the optimization progresses in an
online machine learning approach. In doing so, geometric eects such as diusion
limitations and bifunctional site coupling are accurately captured within the structure
optimization. The impact on the optimal structure is analyzed, yielding new insights
into catalyst structure/activity relationships.
Description
Keywords
Applied sciences, Catalysis, Kinetic Monte Carlo, Sensitivity analysis, Uncertainty quantification