Abstract: Accurate modeling of physical systems is of great importance to engineers. In this thesis, we construct novel machine learning models based on the Neural ODE approach in Chen et al. (2018) and compare them to existing architectures. The evaluation is performed on multiple physical systems, including pendulums and aircraft. Additionally, we investigate the “adjoint method” training algorithm and compare it to backpropagation.
We find that there is usually no benefit to using the “adjoint method” instead of back-propagation due to its slow performance. Finally, we train ODE-models directly from
trajectory data, without access to time derivatives. The proposed models outperform traditional time series models like LSTMs on all investigated datasets.
Github repository here