#NN/ML #sensor #sensing #flexible-structures #machine-learning #3d

Advances in flexible, deployable, and deformable structures and sensors require efficient simulation tools that capture nonlinear geometry and material behavior. We propose a machine learning (ML) approach using neural networks (NN) to simplify simulations, enabling the creation of digital twins and facilitating sim-to-real transfer in structural mechanics.

 

This talk presents a case study using neural networks (NN) to create a reduced-order model for the dynamic simulation of a slinky, a popular children’s toy made of a pre-compressed helical spring that can stretch and deform. Instead of simulating the entire 3D structure of the slinky, we use a reduced representation based on the deformation of its helix axis, significantly reducing the degrees of freedom (DOFs). The mechanics of this simplified representation are captured using a neural ordinary differential equation (neural ODE), trained with data from high-resolution 3D simulations. This approach enables faster dynamic simulations while maintaining physical accuracy, and thanks to the physics-based nature of our model trained with neural ODEs, it is highly generalizable—adapting to changes in boundary conditions or external forces without the need for retraining.

 

The second part of the talk introduces DiSMech, an open-source software platform for fast simulations of flexible structures, which was used in the slinky study. DiSMech aims to enable researchers at all levels to explore the mechanics of soft robots and flexible structures/sensors, driving innovation in robotics research and education. Built on a discrete differential geometry (DDG) approach, it offers a practical alternative to computationally intensive conventional simulation tools.