Tailoring interactions between active nematic defects with reinforcement learning

Active nematics are paradigmatic active matter systems which generate micron-scale patterns and flows. Recent advances in optical control over molecular motors now allow experimenters to control the non-equilibrium activity field in space and time and, in turn, the patterns and flows. However, engineering effective activity protocols remains challenging due to the complex dynamics. Here, we explore a model-free approach for controlling active nematic fields using reinforcement learning. Combining machine learning with trial-and-error exploration of the system dynamics, reinforcement learning bypasses the need for accurate parameterization and model representation of the active nematic. We apply this technique to demonstrate how local activity fields can induce effective interactions between nematic defects, enabling them to follow designer dynamical laws. Moreover, the sufficiency of our low-dimensional system observables and actions suggests that coarse projections of the active nematic field can be used for precise feedback control, making experimental or biological implementation of such feedback loops plausible.