Controlling Defects and Probing Dynamics in Active Nematics with Deep Reinforcement Learning

Abstract

Topological defects govern much of the flow behavior and orientational order in active nematics, making their control relevant for active matter physics, smart materials, and microfluidics. Applied activity patterns can induce self-propulsion of active nematic defects, but general-purpose methods for exploiting this effect to control defects remain largely unexplored. Here we use deep reinforcement learning (RL) to perform minimum-time position control of +1/2 defects in hybrid lattice Boltzmann simulations of active nematodynamics. Spatiotemporally patterned activity, implemented as a control field in the active stress, steers defects through microchannel geometries and reveals finite-time reachable regions of defect position space. Reachability is shaped by director anisotropy, homeotropic wall anchoring, and the allowed activity patterns: local patterns steer defects in free domains but fail in junctions, whereas global patterns open otherwise inaccessible channels. In constrained geometries, the original defect may be unable to reach some goals intact, but controlled pair creation enlarges the effective reachable set by transferring control to a newly created +1/2 defect. The trained RL controllers outperform static and rule-based baselines, and controllers trained only on simple junctions can be combined without fine-tuning into a meta-controller that successfully steers defects through a larger test maze. Free energy visualizations show that guided defects write persistent, history-dependent distortions into the director field that can later be partially erased by -1/2 defects. Thus, RL-based control uncovers how confinement, anchoring, actuation geometry, and defect creation determine reachable motion in active nematics, providing a framework for other control tasks in soft and active matter.