Using unsupervised machine learning to detect depinning and clustering transitions in active matter driven across quenched disorder

Using large-scale numerical simulations of active disks, we demonstrate that a machine learning order parameter can detect depinning transitions and different dynamic flow phases in systems driven far from equilibrium. We model active agents as monodisperse disks executing run-and-tumble motion subject to an external driving force in an environment of quenched disorder. The machine learning order parameter identifies a variety of transitions including the formation of clogged states at low applied drive, the depinning transition and onset of clustering states at intermediate drives and laned states at high drives. These phase transitions are not readily distinguished with traditional measurements, such as the average cluster size and giant cluster fluctuations. We develop several order parameters using principal component analysis applied to particle location, local ordering, and local velocity and compare these to traditional measures. Our results should be useful to characterize a broad class of particle-based systems that exhibit depinning and clustering transitions.