Data-efficient optimization of thermally-activated polymer actuators through machine learning

Abstract

For applications in soft robotics and smart textiles, thermally-activated, twisted, and coiled polymer actuators can offer high mechanical actuation with proper optimization of their processing conditions. However, optimization is often aggravated by the potentially high number of processing variables and the time-consuming nature of materials synthesis and characterization. To overcome these problems, we employed an active machine learning workflow using Bayesian optimization. We subsequently used this approach to optimize the actuation of polymer coils based on three common processing conditions consisting of ply number, applied twisting and coiling stresses. Since the experimental parameters are discrete and not continuous as in conventional Bayesian optimization tasks, a discrete Bayesian optimization workflow was developed. An actuation strain of 1.25 was achieved by optimizing the processing conditions, which required the fabrication of only 62 sample combinations out of 1089 possible ones. Our results highlight the potential of Bayesian optimization in actuator design problems, thereby opening up possibilities for tackling more complex challenges by considering a broader range of processing conditions or addressing multi-objective tasks.