Iterative Refinement of Biohybrid Swimming Devices Using Neural Networks

Biohybrid swimmers have recently emerged as a new form of self-powered swimming device that combines living tissue with artificial components. Built at the millimeter scale, these devices have proposed applications in remote sensing, medical procedures, and for studying muscle regeneration and development. In building these devices, researchers often employ a ‘biomimetic’ design strategy, which attempts to recreate the specific shape or design of a naturally occurring marine lifeforms. However, in nature marine lifeforms display a wide range of different fin shapes and muscular structures depending on their local swimming environment. To select from among these various swimming strategies for use in biohybrid swimmers, here we will discuss how machine learning can be used to guide the engineering-design process. Starting from random initial configurations, fin performance can first be approximated using computational fluid dynamics (CFD) simulations. The resulting data can then be used to train a neural network model, which can make informed predictions about how different fin geometries will perform as biohybrid swimmers. Testing these predictions using experimentally realized tissue engineered devices, we then plan to show how machine learning can be used to improve biohybrid performance, resulting in faster swimming velocities at the millimeter length scale. Overall, by using a machine learning directed process we hope to gain a better quantitative understanding of muscular structure-function relationships, while providing a novel tool for designing biohybrid robotic systems.