Addressing materials’ microstructure diversity using transfer learning

• Published in: npj computational materials Vol. 8; no. 1; pp. 1 - 13
• Main Authors: Goetz, Aurèle, Durmaz, Ali Riza, Müller, Martin, Thomas, Akhil, Britz, Dominik, Kerfriden, Pierre, Eberl, Chris
• Format: Journal Article
• Language: English
• Published: London Nature Publishing Group UK 03.02.2022; Nature Publishing Group; Springer Nature; Nature Portfolio
• Subjects: 639/301/1023/1026; 639/301/930/12; 639/301/930/2735; 639/301/930/328/1649; Automation; Bainite; Characterization and Evaluation of Materials; Chemistry and Materials Science; Computational Intelligence; Datasets; Deep learning; Domains; High strength steels; Image segmentation; Materials Science; Mathematical and Computational Engineering; Mathematical and Computational Physics; Mathematical Modeling and Industrial Mathematics; Mechanics; Microstructure; Neural networks; Photomicrographs; Physics; Scanning electron microscopy; Steel; Structural mechanics; Theoretical; Transfer learning
• ISSN: 2057-3960; 2057-3960
• DOI: 10.1038/s41524-022-00703-z

Materials’ microstructures are signatures of their alloying composition and processing history. Automated, quantitative analyses of microstructural constituents were lately accomplished through deep learning approaches. However, their shortcomings are poor data efficiency and domain generalizability across data sets, inherently conflicting the expenses associated with annotating data through experts, and extensive materials diversity. To tackle both, we propose to apply a sub-class of transfer learning methods called unsupervised domain adaptation (UDA). UDA addresses the task of finding domain-invariant features when supplied with annotated source data and unannotated target data, such that performance on the latter is optimized. Exemplarily, this study is conducted on a lath-shaped bainite segmentation task in complex phase steel micrographs. Domains to bridge are selected to be different metallographic specimen preparations and distinct imaging modalities. We show that a state-of-the-art UDA approach substantially fosters the transfer between the investigated domains, underlining this technique’s potential to cope with materials variance.