guest ::
| Home > Publications database > Data-driven methods for accelerating microstructure-property analysis in metallic materials |
| Home > Publications database > Data-driven methods for accelerating microstructure-property analysis in metallic materials |
| Dissertation / PhD Thesis | FZJ-2026-03324 |
; ;
2026
This record in other databases:
Please use a persistent id in citations: doi:10.34734/FZJ-2026-03324
Abstract: Understanding how microstructural features govern macroscopic properties is a central topic in materials science. Conventional experimental methods, although powerful, face limitations in resolving complex spatiotemporal phenomena and in providing sufficient data for quantitative modeling. This thesis addresses these challenges by applying materials data science approaches, including quantitative data mining of in situ TEM experiments and machine learning-based analysis of nanoindentation data, to enable rapid and quantitative materials characterization of metallic materials. Two complementary research questions guided this work: (i) how to extract richer mechanistic insights and extend the interpretability of experimental observations through data-driven methodologies, and (ii) how to determine the sufficiency of experimental data for robust machine learning applications. To this end, two case studies are presented. The first uses in situ TEM to investigate dislocation dynamics in a Cantor alloy. Treating dislocations as probes of the local energy landscape and applying data-mining techniques allows for the quantitative characterization of the strength, evolution, and spatial distribution of pinning points. The results show that pinning strength varies as dislocations pass, that pinning sites can shift position, and that their spatial distribution is heterogeneous. These findings help explain how microstructural obstacles influence plastic deformation and hardening. The second study applies machine learning to analyze nanoindentation data from Cu–Cr composites with controlled heterogeneity. A Gaussian mixture model is used to identify mechanical phases and estimate their volume fractions directly from experimental data. More importantly, a cross-validation framework is introduced to assess how much data is required for stable model performance, providing practical guidance for experimental data collection. Together, these studies illustrate how data-driven analysis can make experimental data more efficient to use and more informative in revealing underlying mechanisms. The methods developed in this thesis strengthen the quantitative link between microstructure and properties and, in doing so, support the ongoing transition of materials science toward a data-driven paradigm with practical routes for accelerated discovery and design.
|
The record appears in these collections: |
PDF
Fulltext