Predictive Modeling of Soft Stretchable Nanocomposites Using Recurrent Neural Networks

Abstract

Human skin is characterized by rough, elastic, and uneven features that are difficult to recreate using conventional manufacturing technologies and rigid materials. The use of soft materials is a promising alternative to produce devices that mimic the tactile capabilities of biological tissues. Although previous studies have revealed the potential of fillers to modify the properties of composite materials, there is still a gap in modeling the conductivity and mechanical properties of these types of materials. While traditional Finite Element approximations can be used, these methodologies tend to be highly demanding of time and processing power. Instead of this approach, a data-driven learning-based approximation strategy can be used to generate prediction models via neural networks. This paper explores the fabrication of flexible nanocomposites using polydimethylsiloxane (PDMS) with different single-walled carbon nanotubes (SWCNTs) loadings (0.5, 1, and 1.5 wt.%). Simple Recurrent Neural Networks (SRNN), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU) models were formulated, trained, and tested to obtain the predictive sequence data of out-of-plane quasistatic mechanical tests. Finally, the model learned is applied to a dynamic system using the Kelvin-Voight model and the phenomenon known as the bouncing ball. The best predictive results were achieved using a nonlinear activation function in the SRNN model implementing two units and 4000 epochs. These results suggest the feasibility of a hybrid approach of analogy-based learning and data-driven learning for the design and computational analysis of soft and stretchable nanocomposite materials.