Machine learning for soft matter image analysis

Abstract

Over the last few years there has been a great increase in interest in the application of machine learning to problems in physics. Within machine learning there is a wide range of methods that can be used for a huge variety of different problems applicable to soft matter such as the classification of data and the processing of images, but the use of these methods in the field is still in its infancy. This thesis aims to explore how well-established machine learning methods can be put to use to gain more information from historical image data of soft matter systems. The data investigated were historical confocal images of particle-stabilised emulsion systems: bicontinuous interfacially jammed emulsion gels (bijels) and emulsion droplets stabilised by oppositely charged particles. The first bijel data were a collection of images taken from a series of successful and failed bijel syntheses. Machine learning classification algorithms were trained on this data to create a tool for predicting whether an unseen image is from a successful or unsuccessful bijel. A variety of techniques were trialled on a range of variables derived from the autocorrelation function and structure factor of the images, and the best-performing combination of these formed the final model for use in classifying future images. Further data was then added to the model in the form of a second imaging channel and the process was repeated: this model had an improved predictive accuracy. It was found that variables associated with a characteristic length scale in the image were the most useful for classification. The second data were confocal images of droplets stabilised by a mixture of positively and negatively charged colloidal particles. Unsupervised machine learning was used to look for trends in the data that could not be found by traditional methods. These traditional methods were also applied to the data to ascertain how droplet roundness changed as the ratio of positive to negative particles increased. Again, a few different algorithms and variable sets were trialled in order to find the most useful combination. I found that the features I considered follow a different trend to that seen for the size of the droplets in the literature. Finally, machine learning was applied to images of the path of air bubbles through bijel samples after centrifugation in order to ascertain the extent of the effect of the bubble’s passage on the bijel structure. I used a regression algorithm to investigate how the distance from the track left by the bubble could be predicted based on the autocorrelation function of sections of the image. The results can be compared to very recent research using bubbles for quantitative micro-rheology. These examples demonstrate a range of machine learning techniques and their application to soft matter image analysis. One output is a bijel classification tool. With this, and by demonstrating a deeper understanding of the system through data exploration, machine learning is shown to have a wide range of potential uses within the field. The examples also highlight a persistent problem created by the small amounts of data usually available, and some strategies for using these data as effectively as possible are demonstrated.