Controlling Molecules, Particles and Liquids with Thermofluidics and Reinforcement Learning

In recent years, novel techniques have emerged for manipulating molecules, particles and liquid flows within a fluidic environment through remote light control. These methods complement the force generation capabilities of optical tweezers and leverage plasmonic structures [1]. By creating dynamic temperature gradients at solid-liquid interfaces through plasmonic structures, localized stresses are generated, enabling the guidance and manipulation of suspended objects without the necessity of external forces [2]. This field of thermoplasmonics offers a range of non-invasive, highly precise tools for nanoscale feedback-controlled manipulations that can even serve as the foundation for constructing dynamical systems capable of performing reservoir computing [3].

In this contribution, we present experimental results on the integration of thermofluidic systems with reinforcement learning algorithms to enable online learning directly from experimental data. The dynamic optical control of temperature fields enables the generation of local thermo-phoretic, thermo-osmotic, or thermo-viscous actions that are adaptively correlated with the current state of the sample, such as the particle distribution. For this purpose, the sample state is analyzed at video rate and input to actor-critic reinforcement learning algorithms. These algorithms are demonstrated to enable the steering of particles toward predefined goals in the presence of additional fluid flows. This approach holds particular promise for more complex suspensions, offering the discovery and understanding of dynamic flow patterns for separating and manipulating species.