Dead or alive: Distinguishing active from passive particles using supervised learning

TL;DR

This study compares machine learning methods to identify active particles in dense mixtures, revealing that structural features can distinguish activity at high levels but not at low, with implications for biological systems.

Contribution

It introduces a new Voronoi tessellation-based method for identifying active particles, which is faster and requires fewer features than existing approaches.

Findings

Both methods accurately identify active particles at high activity levels.

Voronoi method is faster to train and deploy.

Structural signatures differ at low activity, reducing detection accuracy.

Abstract

A longstanding open question in the field of dense disordered matter is how precisely structure and dynamics are related to each other. With the advent of machine learning, it has become possible to agnostically predict the dynamic propensity of a particle in a dense liquid based on its local structural environment. Thus far, however, these machine-learning studies have focused almost exclusively on simple liquids composed of passive particles. Here we consider a mixture of both passive and active (i.e.\ self-propelled) Brownian particles, with the aim to identify the active particles from minimal local structural information. We compare a state-of-the-art machine learning approach for passive systems with a new method we develop based on Voronoi tessellation. Both methods accurately identify the active particles based on their structural properties at high activity and low concentrations of active particles. Our Voronoi method is, however, substantially faster to train and deploy because it requires fewer, and easy to compute, input features. Notably, both become ineffective when the activity is low, suggesting a fundamentally different structural signature for dynamic propensity and non-equilibrium activity. Ultimately, these efforts might also find relevance in the context of biological active glasses such as confluent cell layers, where subtle changes in the microstructure can hint at pathological changes in cell dynamics.