Tracking bacterial active matter in complex environments using deep learning

Active matter refers to materials that consist of energy-harvesting sub-components that interact to give non-trivial sub- to super-diffusive properties. Tracking such components is challenging due to their variable size, shape, and density—especially in complex environments—making traditional, hand-tuned algorithms unreliable, which limits their generality. We address this by applying deep learning to perform single-particle tracking of bacterial active matter in environments with walls and obstructions by chaining two neural networks for simultaneous detection and trajectory linking that are subsequently analyzed using standard statistical techniques. Remarkably, our models require only a single training example to generalize, enabling the detection of near-bulk flow in densely interacting systems. This approach obtains data that bridge the gap between discrete and continuous descriptions of active matter. It allows for a scalable path to test established theories against experimental data and hence uncover the fundamental principles that drive collective behavior in active matter.