$Φ$-DVAE: Learning Physically Interpretable Representations with Nonlinear Filtering.