Navigation of smart artificial microswimmers in confined flow via reinforcement learning

Artificial microswimmers propel like natural micro-organisms by breaking symmetry and employing unique physico-chemical mechanisms in real environments to perform complex tasks. Recent studies with 'smart' microswimmers have expanded the scope with respect to their responses to external/tactic stimuli, hydrodynamic traps and multiple functions such as aggregated swimming. However, the rapid and precise regulation of these smart swimmers mediated by imposed flow conditions remains elusive. Here, we model them as active Brownian particles which satisfy a system of overdamped stochastic Langevin equations of motion that determine their position and orientation in a pressure-driven confined flow. To attain precise navigation from random position to pre-defined target, we intermittently guide these particles via reinforcement learning wherein the numerical rewards depend on their choice of actions that is derived from a discrete set of possible orientations. Further, to mitigate the reliance on domain knowledge, we automate the reward policy based on a desired sequence of actions which has the potential to eventually bring these smart microswimmers closer to nature.