Tracking Janus microswimmers in 3D with Machine Learning

Abstract

Advancements in artificial active matter heavily rely on our ability to characterise their motion. Yet, the most widely used tool to analyse the latter is standard wide-field microscopy, which is largely limited to the study of two-dimensional motion. In contrast, real-world applications often require the navigation of complex three-dimensional environments. Here, we present a Machine Learning (ML) approach to track Janus microswimmers in three dimensions, using Z-stacks as labelled training data. We demonstrate several examples of ML algorithms using freely available and well-documented software, and find that an ensemble decision tree-based model (Extremely Randomised Decision Trees) performs the best at tracking the particles over a volume spanning a depth of more than 40 $\mu$m. With this model, we are able to localise Janus particles with a significant optical asymmetry from standard wide-field microscopy images, bypassing the need for specialised equipment and expertise such as that required for digital holographic microscopy. We expect that ML algorithms will become increasingly prevalent by necessity in the study of active matter systems, and encourage experimentalists to take advantage of this powerful tool to address the various challenges within the field.

40 pages, 10 figures

References in corpus (8)

• Self-motile colloidal particles: from directed propulsion to random walk
• Propulsion of a molecular machine by asymmetric distribution of reaction--products
• Induced-charge Electrophoresis of Metallo-dielectric Particles
• Phoretic Motion of Spheroidal Particles Due To Self-Generated Solute Gradients
• Active phase separation by turning toward regions of higher density
• Quantitative Digital Microscopy with Deep Learning
• Diffusion-based height analysis reveals robust microswimmer-wall separation
• Microrobots from Toposelective Nanoparticle Attachment

Cited by in corpus (4)

• Colloidal Hard Spheres: Triumphs, Challenges and Mysteries
• Discovering dynamic laws from observations: the case of self-propelled, interacting colloids
• Janus microswimmers are poor hydrodynamic mixers
• Convolutional neural networks applied to differential dynamic microscopy reduces noise when quantifying heterogeneous dynamics