Data Driven Long-Term Energy Efficiency Prediction of Dielectric Elastomer Artificial Muscles

Dielectric elastomer actuators (DEAs) are electro-mechanical transducers driven by electric fields. They can be made from all solid-state materials, allowing them to operate under extreme environments, be scalable in size, and be capable of self-sensing. Their potential applications span robotics, haptics, and optical instruments. However, practical adoption is hindered by the absence of a material design framework and a non-destructive performance evaluation system. This research introduces a data-driven framework to predict long-term DEA energy efficiency using short-term electrical property measurements. The DEA datasets are generated from single-layer pre-stretched actuators made with different electrode and elastomer materials, being tested for 30 minutes using a custom-made testing instrument. First, experiments were conducted to develop an empirical understanding of the electro-mechanical energy conversion mechanism and the impact of material choices during DEA actuation. Second, data-driven models were applied to the datasets to predict energy efficiency at 30 minutes using as short as 1 minute of electromechanical measurements. DEAs Third, the potential of generalizing this framework was investigated by adopting transfer learning to predict the 3-hour energy efficiency of single-layer DEAs using as short as 1-minute input data and to extend the predictions on devices made of a multi-layer architecture. Transfer learning is an ML strategy that fine-tunes a pre-trained model on a larger dataset, then adapts it to a smaller, high-value dataset with limited samples, leveraging knowledge transfer to enhance prediction performance in data-scarce scenarios. Despite significant differences in behavior between single-layer and multi-layer DEAs, transfer learning bridges the gap in performance predictability, highlighting the need for data-driven approaches to performance prediction. We believe the proposed data-driven framework can evolve into an intelligent system for accelerating new material discovery by predicting performance and providing device optimization strategies.