Physics-Based Interpolation For Animation Using Physics-Informed Neural Networks
Author
Metadata
Show full item recordAbstract
Keyframe interpolation is a fundamental technique used in computer animation. In this technique, an animator specifies keyframes and has a computer program generate the in-betweens by interpolating between them. However, most common forms of interpolation use a non-physics-based path between keyframes to decide where the in-betweens should be drawn. When animating physical behavior, interpolating along a physics-based path of motion could create more realistic animations while reducing the number of keyframes that must be specified by the animator. Physically Informed Neural Networks (PINNs) are neural networks used to solve problems involving partial differential equations by directly incorporating information about those equations into the network. Because prior knowledge of the functions they are trying to model is incorporated into the network, one advantage of PINNs is they require less data to train than neural networks trained using only data points. This research trains PINNs to interpolate the path of an animated object, specifically a mass attached to a two-dimensional spring, in a physics-based manner while closely meeting the constraints imposed on the object’s position and/or velocity at keyframes. In addition, this project explores how modifying the hyperparameters and boundary conditions used to train the PINN affects the interpolation and the efficiency of the training process
Subject
PINNPhysics Informed Neural Network
Graphics
Computer Graphics
Animation
Keyframe Animation
Keyframe Interpolation
Physics Based Animation
PyTorch
Citation
Alvarez Del Castillo, Carlos (2023). Physics-Based Interpolation For Animation Using Physics-Informed Neural Networks. Undergraduate Research Scholars Program. Available electronically from https : / /hdl .handle .net /1969 .1 /200260.