Dissipative Hamiltonian Neural Networks

Andrew Sosanya, Sam Greydanus | 2020

• Paper: https://arxiv.org/abs/2201.10085
• Blog: greydanus.github.io/2022/01/25/dissipative-hnns/

Dissipative HNNs (D-HNNs) output two scalar functions, denoted here by H and D. The first of these two, H, is the Hamiltonian. It is perfectly conserved. The second of these two, D, is the Rayleigh dissipation function. It models the dissipative component of the dynamics of a physical system. The addition of the dissipation function is what sets this model apart from Hamiltonian Neural Networks; it allows D-HNNs to model systems where energy is not quite conserved, as, for example, in the case of a damped mass-spring system.

Basic usage

Use the .ipnyb notebooks to train and analyze all models

Summary

We propose a simple way of extending Hamiltonian Neural Networks so as to model physical systems with dissipative forces. We call this model a Dissipative Hamiltonian Neural Network (D-HNN) because it adds support for dissipative dynamics.

Dependencies

• PyTorch
• NumPy
• ImageIO
• Scipy

This project is written in Python 3.