Edinburgh Research Explorer

Autonomous analysis to identify bijels from two-dimensional images

Research output: Contribution to journalArticle

Related Edinburgh Organisations

Open Access permissions

Open

Documents

  • Download as Adobe PDF

    Accepted author manuscript, 3.13 MB, PDF document

    Licence: Creative Commons: Attribution (CC-BY)

  • Download as Adobe PDF

    Final published version, 3.46 MB, PDF document

    Licence: Creative Commons: Attribution (CC-BY)

Original languageEnglish
JournalSoft Matter
DOIs
Publication statusPublished - 11 Feb 2020

Abstract

Bicontinuous interfacially jammed emulsion gels (bijels) are novel composite materials that can be challenging to manufacture. As a step towards automating production, we have developed a machine learning tool to classify fabrication attempts. We use training and testing data in the form of confocal images from both successful and unsuccessful attempts at bijel fabrication. We then apply machine learning techniques to this data in order to classify whether an image is a bijel or a non-bijel. Our principal approach is to process the images to find their autocorrelation function and structure factor, and from these functions we identify variables that can be used for training a supervised machine learning model to identify a bijel image. We are able to categorise images with reasonable accuracies of 85.4% and 87.5% for two different approaches. We find
that using both the liquid and particle channels helps to achieve optimal performance and that successful classification relies on the bijel samples sharing a characteristic length scale. Our second approach is to classify the shapes of the liquid domains directly; the shape descriptors are
then used to classify fabrication attempts via a decision tree. We have used an adaptive design approach to find an image pre-processing step that yields the optimal classification results. Again, we find that the characteristic length scale of the images is crucial in performing the classification.

Download statistics

No data available

ID: 134990560