Can physics help predict hidden parts of chaotic systems?

Learning dynamics from partial observations is hard: you see some variables but not others, yet standard physics-informed models assume full state access. This work combines Hamiltonian neural networks with neural ODEs to learn from incomplete data, building energy conservation directly into the model. Tested on mass-spring systems and the three-body problem, the method outperforms data-only baselines on long-horizon predictions, staying stable where competitors diverge.