A problem specific recurrent neural network for the description and simulation of dynamic spring models

We present a recurrent neural network which was designed for the description and simulation of dynamic spring models. The network simulates the physical behavior of deformable or elastic solids like stiffness, viscosity and inertia. The physical parameters of the real model can be used to initialize the network parameters. Besides, it is possible to learn the deformation behavior of a real solid. Using a neural network structure, local changes to the system like collisions or cuts can be easily performed during simulation. Furthermore, it is possible to speed up the simulation by parallel hardware.