The Department of Civil and Systems Engineering
and
Advisor Lori Graham-Brady, Professor
Announce the Thesis Defense of
Noah Wade
Monday, March 28
9:00 am
Latrobe 106
QUANTIFYING MEASUREMENT UNCERTAINTIES USING DIGITAL REPRESENTATIONS OF MATERIALS
Material characterization is important in many different engineering disciplines. It provides valuable information to computational models with far reaching applications, but it is also challenging and highly specialized. As such, there is potential to apply undue confidence in characterization data as it is collected and passed to modelers for subsequent analysis.
This thesis looks at uncertainty quantification in material characterization at this systems level. It considers not only the errors that might be present in a given measurement but also how those errors propagate into subsequent computational models. An uncertainly quantification framework uses a simulation-based approach to identify, measure, and propagate sources of uncertainty that arise in typical image-based material characterization workflows. The framework uses the concept of a virtual material to estimate probabilistic characteristics of measurement error. The framework is applied to various characteristics to show how different sources of uncertainty propagate through the system and suggest how characterization can be made more efficient and accurate.
The first part of this work applies this framework to the process of characterizing a 3D crystalline microstructure. Simulations of electron backscatter diffraction provide a virtual data set that is subsequently evaluated using finite element simulation. Results show that some experimental efforts to reduce uncertainties prove unnecessary, since those sources of error do not propagate through the system. At the same time a number of other sources of error led to significant uncertainties in the end product computational models. Sampling resolution emerged as a source of error that has significant influence on the subsequent analyses. The second part of this work further examines the effects of sampling resolution. Simulated data collection provides a conditional distribution on key measurements as a function of resolution, from which a Bayesian approach yields a distribution of the true material characteristic given a particular measurement of that characteristic. These conditional measurement distributions allow for more efficient and effective collection of microstructural data. The third part of this work considers the process of simulating material microstructures with irregular features that are not easily simulated with standard approaches. In particular, this work studies some of the potential sources of error in a machine learning model that is used to reconstruct images based on transfer learning from the VGG-19 network.