Publication: Automatisierte Anbindung von Simulations- und Optimierungssoftware zur parallelen Lösung inverser Problemklassen
Introduction
Applications
Tools
Research Groups
Workshops
Publications
   List Publications
   Advanced Search
   Info
   Add Publications
My Account
About

Automatisierte Anbindung von Simulations- und Optimierungssoftware zur parallelen Lösung inverser Problemklassen

- Ph.D. thesis -
 

Area
Inverse Problems, Optimization

Author(s)
R. Seidler

Year
2020

Abstract
This work introduces a novel extension of the Python software Environment for Combining Optimization and Simulation Software (EFCOSS). The extension addresses the solution of optimization problems of different types. Various problem instances that demonstrate the feasibility of this approach in new practical application scenarios include geothermal engineering and material science. In a more general context, EFCOSS enables to investigate the questions of which parameter values best fit a given computer model to measurements from real-world experiments, how should such experiments be designed with minimal uncertainty, which computer models should be used for a specific task, and how can the efficiency of such investigations be improved by using simpler models. These questions are addressed by employing techniques from parameter estimation, space mapping, optimal experimental design, and model identification that are implemented and brought together in EFCOSS. To this end, the internal structure of EFCOSS had to be redesigned completely to introduce a new software architecture based on standard Python packages. This new architecture allows for multiple optimization problems, objective functions, and simulation routines to be used within a single application. Simulation models of geothermal reservoirs in the regions of Perth in Australia and Tuscany in Italy are used as illustrating examples of optimal experimental design to find the location of new borehole sites that introduce low uncertainty in the parameter estimation. Furthermore, new parameter estimation problems are solved using space-mapping algorithms. Model identification is applied to metal-plasticity models to investigate different kinds of models. By combining automatic differentiation, parallelization, and reuse of intermediate results, the serial runtimes of the described problems are reduced from several weeks to minutes.

AD Tools
TAPENADE

BibTeX
@PHDTHESIS{
         Seidler2020AAv,
       url = "https://doi.org/10.22032/dbt.46968",
       doi = "10.22032/dbt.46968",
       abstract = "This work introduces a novel extension of the Python software
         \emph{Environment for Combining Optimization and Simulation Software} (EFCOSS). The extension
         addresses the solution of optimization problems of different types. Various problem instances that
         demonstrate the feasibility of this approach in new practical application scenarios include
         geothermal engineering and material science. In a more general context, EFCOSS enables to
         investigate the questions of which parameter values best fit a given computer model to measurements
         from real-world experiments, how should such experiments be designed with minimal uncertainty, which
         computer models should be used for a specific task, and how can the efficiency of such
         investigations be improved by using simpler models. These questions are addressed by employing
         techniques from parameter estimation, space mapping, optimal experimental design, and model
         identification that are implemented and brought together in EFCOSS. To this end, the internal
         structure of EFCOSS had to be redesigned completely to introduce a new software architecture based
         on standard Python packages. This new architecture allows for multiple optimization problems,
         objective functions, and simulation routines to be used within a single application. Simulation
         models of geothermal reservoirs in the regions of Perth in Australia and Tuscany in Italy are used
         as illustrating examples of optimal experimental design to find the location of new borehole sites
         that introduce low uncertainty in the parameter estimation. Furthermore, new parameter estimation
         problems are solved using space-mapping algorithms. Model identification is applied to
         metal-plasticity models to investigate different kinds of models. By combining automatic
         differentiation, parallelization, and reuse of intermediate results, the serial runtimes of the
         described problems are reduced from several weeks to minutes.",
       author = "R. Seidler",
       title = "Automatisierte {A}nbindung von {S}imulations- und {O}ptimierungssoftware zur
         parallelen {L}{\"o}sung inverser Problemklassen",
       school = "Department of Mathematics and Computer Science, Friedrich Schiller University Jena",
       year = "2020",
       type = "Dissertation",
       address = "Jena",
       ad_area = "Inverse Problems, Optimization",
       ad_tools = "TAPENADE"
}


back
  

Contact:
autodiff.org
Username:
Password:
(lost password)