Exploiting the Dynamics of Soft Materials for Machine Learning

TLDR

Soft materials are increasingly used in engineering, enabling functions difficult with rigid materials and offering a radical new way to compute by exploiting their physical dynamics. The authors propose that the diverse dynamics generated by actuating soft materials can serve as a resource for machine learning. They demonstrate this by multiplexing the transient dynamics of a soft silicone arm and evaluating its performance on two standard benchmark tasks. The soft arm performs comparably to or better than conventional machine learning methods across multiple conditions and can predict its own sensory time series, showing immediate real‑world applicability.

Abstract

Soft materials are increasingly utilized for various purposes in many engineering applications. These materials have been shown to perform a number of functions that were previously difficult to implement using rigid materials. Here, we argue that the diverse dynamics generated by actuating soft materials can be effectively used for machine learning purposes. This is demonstrated using a soft silicone arm through a technique of multiplexing, which enables the rich transient dynamics of the soft materials to be fully exploited as a computational resource. The computational performance of the soft silicone arm is examined through two standard benchmark tasks. Results show that the soft arm compares well to or even outperforms conventional machine learning techniques under multiple conditions. We then demonstrate that this system can be used for the sensory time series prediction problem for the soft arm itself, which suggests its immediate applicability to a real-world machine learning problem. Our approach, on the one hand, represents a radical departure from traditional computational methods, whereas on the other hand, it fits nicely into a more general perspective of computation by way of exploiting the properties of physical materials in the real world.