Predictive Modeling of Soft Stretchable Nanocomposites Using Recurrent Neural Networks.

Area of Science:

• Materials Science
• Computational Modeling
• Soft Robotics

Background:

• Mimicking human skin's tactile properties requires advanced materials beyond conventional rigid manufacturing.
• Soft composite materials offer a promising avenue for creating bio-inspired devices.
• Accurate modeling of soft composite properties is challenging with traditional methods.

Purpose of the Study:

• To fabricate flexible nanocomposites using polydimethylsiloxane (PDMS) and single-walled carbon nanotubes (SWCNTs).
• To develop and evaluate data-driven neural network models for predicting mechanical properties.
• To apply learned models to dynamic systems for computational analysis.

Main Methods:

• Fabrication of PDMS/SWCNT nanocomposites with varying SWCNT concentrations (0.5, 1, 1.5 wt.%).
• Development and testing of Simple Recurrent Neural Networks (SRNN), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU) for property prediction.
• Application of trained models to Kelvin-Voight dynamic systems, including the bouncing ball phenomenon.

Main Results:

• Neural network models, particularly SRNN with a nonlinear activation function, accurately predicted material behavior.
• The SRNN model with two units and 4000 epochs yielded the best predictive performance.
• Successful application of the learned model to simulate dynamic mechanical responses.

Conclusions:

• Data-driven learning offers an efficient alternative to traditional methods for modeling soft nanocomposites.
• A hybrid approach combining analogy-based and data-driven learning is feasible for designing and analyzing these materials.
• This research paves the way for advanced soft and stretchable material development.