Fluid intelligence: activity from learning and forgetting

Active systems ranging from cellular colonies to collections of nanomotors typically derive their unusual hydrodynamic properties from the exertion of mechanical work at small scales. However, when the microscopic constituents can record measurements of their environment, thermodynamics becomes inextricably intertwined with information processing. To single-out how information processing affects the dynamics of active fluids, we propose a model system in which the microscopic constituents reach non-equilibrium steady states by taking measurements of their surroundings and performing zero-work actions which exploit that information. Using molecular dynamics simulations, kinetic theory and hydrodynamics we explore the properties of this class of many-body Maxwell-demon systems. Furthermore, we employ reinforcement learning to explore goal-oriented policies of constituent behavior and find that this form of information-based activity is robust even when the agitating environment is athermal. Our model system is illustrative of a broader class of active systems in which intelligent actors choose low-effort actions that exploit environmental fluctuations to achieve nonequilibrium states.