Assessing the properties of a colloidal suspension with the aid of deep learning

TL;DR

This study demonstrates that convolutional neural networks can effectively classify complex colloidal suspensions from speckle images, outperforming traditional methods in accuracy and efficiency, with promising but limited generalization capabilities.

Contribution

The paper introduces a CNN-based approach for classifying colloidal suspensions from speckle images, showing high accuracy across many classes and comparing it with SVM methods.

Findings

CNN accurately classifies 73 suspension types.

SVM is more resource-intensive and less scalable.

Image fragmenting reduces classification accuracy.

Abstract

Convolution neural networks were applied to classify speckle images generated from nano-particle suspensions and thus to recognise suspensions. The speckle images in the form of movies were obtained from suspensions placed in a thin cuvette. The classifier was trained, validated and tested on both single component monodispersive suspensions, as well as on two-component suspensions. It was able to properly recognise all the 73 classes - different suspensions from the training set, which is far beyond the capabilities of the human experimenter, and shows the capability of learning many more. The classes comprised different nanoparticle material and size, as well as different concentrations of the suspended phase. We also examined the capability of the system to generalise, by testing a system trained on single-component suspensions with two-component suspensions. The capability to generalise was found promising but significantly limited. A classification system using neural network was also compared with the one using support vector machine (SVM). SVM was found much more resource-consuming and thus could not be tested on full-size speckle images. Using image fragments very significantly deteriorates results for both SVM and neural networks. We showed that nanoparticle (colloidal) suspensions comprising even a large multi-parameter set of classes can be quickly identified using speckle images classified with convolution neural network.