Speaker
Description
Natural microswimmers move in low Reynolds number environments by performing non-reciprocal body deformations that generate propulsion in viscous fluids. Inspired by these mechanisms, artificial microswimmers are being designed to perform tasks such as targeted transport and drug delivery.
In this work, we train a two-dimensional triangular swimmer to move in a desired direction using different propulsion gaits. The swimmer’s dynamics are controlled by adaptive neural networks that map its internal degrees of freedom to applied forces. These networks are optimized using the NEAT (NeuroEvolution of Augmenting Topologies) algorithm, which evolves both the weights and the architecture of the network.
Extending earlier studies [1] on one-dimensional three-bead swimmers navigating chemical landscapes, we consider the more complex case of a two-dimensional triangular swimmer [2]. The system exhibits several emergent non-reciprocal propulsion modes, including flapping, chiral, and walking motions. In two dimensions, a simple reward function based solely on displacement is insufficient to induce effective motion; without a suitable reward scheme, the swimmer tends to remain stationary or move in circular trajectories without net translation. Therefore, a more sophisticated reward function is introduced [2], incorporating instantaneous and average displacement, angular displacement, and shape factors. Analysis of the resulting neural architectures reveals that the emergent networks are relatively simple yet exhibit distinct structural differences corresponding to the various swimming gaits.
Ref.
1. B. Hartl, M. Hϋbl, G. Kahl, and A. Zӧttl, Proc. Natl. Acad. Sci. USA., 118 (2021), e2019683118.
2. R. Maity, M. Hϋbl, J. Lemmel, B. Hartl, and G. Kahl, Emergent swimming strategies of an intelligent triangular three-bead swimmer, Submitted.