Detecting active clustering on quenched disorder with unsupervised machine learning

Phase transitions can be difficult to characterize in active matter due to the soft interparticle interactions and the inherent disorder. Using large-scale numerical simulations of active disks driven far from equilibrium, we demonstrate that principal component analysis, a dimensionality reduction technique popular in machine learning, can detect dynamical phase transitions in systems driven far from equilibrium. We model active agents as monodisperse disks executing run-and-tumble motion in two regimes - motility induced phase transitions in a clean environment, and subject to an external driving force across an environment of quenched disorder. The machine learning order parameter is derived from a system wide measure of interparticle distance, that distinguishes homogeneous versus inhomogeneous arrangements of particles. Using this tool, we identify a variety of order-disorder transitions such as clustering and depinning and disorder-disorder transitions including clogging and laning which are not readily distinguished with traditional measurements. We highlight particularly the detection of incipient clustering, where the machine learning order parameter performs better than measurements based on interparticle contact.