Phase Space Data-Driven Simulation of Elastic Materials

Abstract

The use of data driven models in computer animation offers several benefits. We present an analysis of a regression model as a method to simulate cloth. In our approach, we generate data from a simple mass-spring system and we fit a regressor. Then, we assemble more complex mass-spring systems and use the learnt model to simulate them. To validate the approach we perform several tests. We analyze the elastic properties of a single learnt spring, measuring its stiffness coefficient, and compare it to the original, physics-based, model. We also build several test scenarios which include the simulation of a piece of cloth under gravity, comparing the regression model and the physics-based model. Finally we test the behaviour of the regression model for systems with high stiffness coefficient and compare its stability properties with a semi-implicit Euler integration method.