Joint learning of latent reduced-order dynamics and states from noisy observations, by auto-differentiating through an Ensemble Kalman Filter (EnKF) using PyTorch. This repo is built upon AD-EnKF (https://github.com/ymchen0/torchEnKF).
Compared to AD-EnKF, we enable nonlinear observation model with spectral convolutional layers (see Sec 4.1 of paper). Similar parameterization can also be used for the dynamics model in AD-EnKF to increase flexibility (see Appendix A). The implementation can be found in torchEnKF/nn_templates.py.
(More examples coming soon...)