Reproduce “PINN-Ray: A Physics-Informed Neural Network to Model Soft Robotic Fin Ray Fingers” on OpenResearch

About this paper

Xing Wang, Joel Janek Dabrowski, Josh Pinskier, Lois Liow, Vinoth Viswanathan, Richard Scalzo, David Howard

This research from CSIRO Data61 introduces PINN-Ray, a Physics-Informed Neural Network designed to model the complex, nonlinear deformation of soft robotic Fin Ray fingers. By assimilating just nine experimental data points, PINN-Ray achieves a Mean Absolute Error of 0.03 mm in displacement prediction, significantly outperforming traditional Finite Element Models and standard PINN approaches while also accurately estimating internal stress and strain distributions.

Key insights

• Integrating a small amount of high-fidelity experimental data (9 points) via data assimilation into a physics-informed neural network (PINN) drastically improves model accuracy and closes the 'sim-to-real gap' for soft robotic deformation.
• PINN-Ray, with data assimilation, achieves superior accuracy in predicting both observable displacements and unobservable latent variables like stress and strain compared to traditional Finite Element Models (FEM) and PINNs without assimilation.
• Embedding the principle of minimum potential energy from elastic mechanics directly into the PINN's loss function, combined with automatic differentiation, enables the network to learn physically consistent deformation fields.

Summary by alphaXiv.