% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@PHDTHESIS{Heymann:996812,
author = {Heymann, William},
othercontributors = {Wiechert, Wolfgang and Jupke, Andreas},
title = {{A}dvanced parameter estimation, {B}ayesian uncertainty
quantification, and surrogate modeling for chromatography
processes},
school = {Rheinisch-Westfälische Technische Hochschule Aachen},
type = {Dissertation},
address = {Aachen},
publisher = {RWTH Aachen University},
reportid = {RWTH-2024-10846},
pages = {1 Online-Ressource : Illustrationen},
year = {2023},
note = {Veröffentlicht auf dem Publikationsserver der RWTH Aachen
University 2024; Dissertation, Rheinisch-Westfälische
Technische Hochschule Aachen, 2023},
abstract = {Manufacturing medicine is hard and expensive. Manufacturing
protein-based medicines can be especially difficult due to
the size and complexity of the molecules. Packed bed liquid
chromatography is used during the manufacturing process to
remove impurities. The problem with liquid phase
chromatography is that it is an expensive and difficult
process to develop. The goal of chromatography modeling is
to make it faster and easier to develop a chromatography
process that purifies that target protein. Chromatography
modeling is difficult and the process of designing a model,
designing experiments, calibrating the model, and
determining the uncertainty of the model parameters is a
complex and computationally intensive process. This thesis
covers model calibration, parameter uncertainty, and
surrogate modeling. Model calibration is done with parameter
estimation where the parameters of a model are estimated
based on chromatogram data. While least squares estimation
of unknown parameters is a well-established standard it can
also suffer from critical disadvantages. The description of
real-world systems is generally prone to unaccounted
mechanisms in the models that are customarily applied, such
as dispersion in external holdup volumes, and systematic
measurement errors, such as caused by pump delays. In this
scenario, matching the shape between simulated and measured
chromatograms has been found to be more important than the
exact peak positions. A new score system is demonstrated
that separately accounts for the shape, position, and height
of individual peaks. A genetic algorithm is used for
optimizing these multiple objectives. Even for
non-conflicting objectives, this approach shows superior
convergence in comparison to single-objective gradient
search, while conflicting objectives indicate incomplete
models or inconsistent data. In the latter case, Pareto
optima provide important information for understanding the
system and improving experiments. Once a model is calibrated
the next step is to determine how well defined the model is.
With increased dependence on modeling, it is not enough to
calibrate a model to experimental data. No model is perfect,
and no model can perfectly explain experimental data. What
is increasingly needed is the uncertainty in the calibrated
model’s parameters along with the uncertainty in the
resulting chromatogram. A method is presented to determine
the probability distribution of parameters for a calibrated
model using Bayesian uncertainty quantification. This method
incorporates experimental errors such as pump delays, pump
flow rates, loading concentration, and UV measurement error.
The method presented here is based on Bayes’ theorem and
uses Markov Chain Monte Carlo with an ensemble sampler and
covers how to build and evaluate the error model. While the
model calibration and parameter uncertainty analysis
presented here work well for synthetic and industrial data,
they also require a lot of computing resources. A surrogate
model approximates the original model and can stand in for
the real model under restricted conditions. The advantage of
surrogate models is they can be tens of thousands of times
faster than the original model. Construction of a surrogate
model using an artificial neural network is demonstrated
along with the entire network design process. All software
created for this project is freely available as open-source
code on GitHub (https://github.com/modsim/CADET-Match).},
cin = {420410},
ddc = {620},
cid = {$I:(DE-82)420410_20140620$},
typ = {PUB:(DE-HGF)11},
doi = {10.18154/RWTH-2024-10846},
url = {https://publications.rwth-aachen.de/record/996812},
}
Gast ::