Microswimmers learning chemotaxis via genetic algorithms

Microorganisms which are abundant in nature play a key role in many biological phenomena. Over ages they have developed a huge variety of strategies which help them in nutrient uptake, to reproduce, to escape from predators, or to hunt prey. These strategies are realized via different manners, such as shape deformation or by the use of appendages. In an effort to understand the locomotion of these microswimmers under low Reynolds number conditions, models have been proposed, which are based on fixed swimming strategies. However, models of such organisms that adapt their swimming strategies to the ever-changing surroundings have, so far, rarely been investigated in literature. It is the aim of this project to study models which are trained to swim and to uptake nutrients with the help of Machine Learning tools. In this project I plan to train simple models of microswimmers (such as the tetrahedral four-bead swimmer, the chiral squirmer, or the three-sphere mirror-symmetric swimmer) to adapt their swimming strategies to their surrounding and to train them in this way to perform specific tasks: learning to swim, nutrition uptake, competition with other microswimmers, etc. To be more specific I will combine the equations-of-motion of the microswimmers with modern tools of Machine Learning in such a manner that the microorganism optimize their migration strategies in an effort to realize well- defined tasks. I will explore these microorganisms in different settings, i.e., different nutrition sources, competition between swimmers in their nutrition uptake, and small ensembles of swimmers. By analyzing the behaviour of the swimmers that they develop in different surroundings we will obtain a deeper insight into their strategies. I want to address this problem by combining conceptual tools of Machine Learning with the formalism that governs the motion of microswimmers at low Reynolds numbers, a route that has hardly been explored so far in literature to tackle this problem. To be more specific I will use adaptive neural networks that control the shape deformation of the models based on the interaction with their surroundings. Thus I will be able to train the swimming gaits of the swimmer to provide directional movements in static and time-dependent chemical environments. The above-mentioned methodological ansatz has been exploited so far only in a first, successful application to a simple, linear three-bead swimmer. The extension of this approach to considerably more complex microswimmer models and to rather complex setups is highly originally and dwells both on my previous expertise accumulated for microswimmers as well as on the experience available in my host group at the Institute for Theoretical Physics at the TU Wien, coordinated by Prof. Gerhard Kahl.