Self-supervised hamiltonian mechanics neural networks

Abstract

We developed a method to derive the hamiltonian of an unknown system by machine learning the motions of the system. We modified the training process of Greydanus et al.’s hamiltonian neural network to make it be capable of learning from a dataset without the change-of-state ground truth. In other word, the learning process is self-supervised. This improvement can extend the application of the hamiltonian neural network because it is sometimes difficult to accurately measure the change of state of the system. Our model can now be able to learn the free particle system and the harmonic oscillator system.