Title:
Data-Driven Mixed-Integer Optimization for Modular Process Intensification
Data-Driven Mixed-Integer Optimization for Modular Process Intensification
Author(s)
Kim, Sophie
Advisor(s)
Boukouvala, Fani
Editor(s)
Collections
Supplementary to
Permanent Link
Abstract
High-fidelity computer simulations provide accurate information on complex physical systems. These often involve proprietary codes, if-then operators, or numerical integrators to describe phenomena that cannot be explicitly captured by physics-based algebraic equations. Consequently, the derivatives of the model are either absent or too complicated to compute; thus, the system cannot be directly optimized using derivative-based optimization solvers. Such problems are known as “black-box” systems since the constraints and the objective of the problem cannot be obtained as closed-form equations. One promising approach to optimize black-box systems is surrogate-based optimization. Surrogate-based optimization uses simulation data to construct low-fidelity approximation models. These models are optimized to find an optimal solution.
We study several strategies for surrogate-based optimization for nonlinear and mixed-integer nonlinear black-box problems. First, we explore several types of surrogate models, ranging from simple subset selection for regression models to highly complex machine learning models. Second, we propose a novel surrogate-based optimization algorithm for black-box mixed-integer nonlinear programming problems. The algorithm systematically employs data-preprocessing techniques, surrogate model fitting, and optimization-based adaptive sampling to efficiently locate the optimal solution. Finally, a case study on modular carbon capture is presented. Simultaneous process optimization and adsorbent selection are performed to determine the optimal module design. An economic analysis is presented to determine the feasibility of a proposed modular facility.
Sponsor
Date Issued
2020-12-06
Extent
Resource Type
Text
Resource Subtype
Dissertation