A reinforcement learning approach to optical control of active matter

Even in bulk, uniform systems, active matter has already shown exciting new physics and potential for materials development. Additionally, spatio-temporal control of active systems is becoming experimentally possible in a variety of ways. Activity itself is an appealing control knob that is qualitatively similar across active matter systems; while active systems can differ in their details, they generally contain a scalar field associated with local microscopic energy dissipation.

However, leveraging high-dimensional spatio-temporal activity patterns is difficult; a brute-force search in such a high-dimensional space is unlikely to be successful, especially without system-specific physical intuition.

Here, we apply a reinforcement learning approach; a reinforcement learning agent identifies time-varying patterns of a scalar activity parameter which induce net transport in a chosen direction in a simulated system of self-propelled spheres. When aligning interactions are added between the spheres, the nature of the patterns learned by the agent changes, illustrating the flexibility of the reinforcement learning approach.