Presentation Information
[PCP2-13]Interactive Microstructure Mapping for Exploring Microstructure–Property Relationships
*Yassin Rahman1, Manuela Erbe1, Kai Walter1, Alexandra Jung1, Bernhard Holzapfel1 (1. Karlsruhe Institute of Technology, Institute of Technical Physics (Germany))
Keywords:
Microstructure Analysis,Machine Learning,Data Analysis,Microstructure-Property-Relationships,Data Mining,Critical Current
SEM images of REBCO thin film layers fabricated via the TFA-MOD process were acquired and compared to investigate structure–property relationships in these superconducting materials.
A software was developed to generate a two-dimensional map where the microstructure images are placed on a 2D map where images with similar visual features are placed close to each other.
To achieve this, openly available convolutional neural networks (CNN) and pretrained weights were used to generate high-dimensional vectors encoding visual features of the images in an abstract way. This vector space is converted into an interactive 2D map by using dimensionality reduction algorithms.
The microstructure images were found to arrange in interpretable ways, and when the target value Jc (critical current density) was colour-encoded on the interactive map, correlations between regions of low Jc and specific types of microstructural features became visible providing a proof-of-concept.
This approach is general and can be applied to arbitrary microscopy images of microstructures. It therefore provides an efficient exploratory tool for detecting links between processing, microstructure and functional properties in a wide range of materials systems.
A software was developed to generate a two-dimensional map where the microstructure images are placed on a 2D map where images with similar visual features are placed close to each other.
To achieve this, openly available convolutional neural networks (CNN) and pretrained weights were used to generate high-dimensional vectors encoding visual features of the images in an abstract way. This vector space is converted into an interactive 2D map by using dimensionality reduction algorithms.
The microstructure images were found to arrange in interpretable ways, and when the target value Jc (critical current density) was colour-encoded on the interactive map, correlations between regions of low Jc and specific types of microstructural features became visible providing a proof-of-concept.
This approach is general and can be applied to arbitrary microscopy images of microstructures. It therefore provides an efficient exploratory tool for detecting links between processing, microstructure and functional properties in a wide range of materials systems.