MOSAIC: Sparse Temporal VAE for Causal Discovery in Scientific Time Series

A new causal representation learning method called MOSAIC (Module discovery via Sparse Additive Identifiable Causal learning) has been proposed for scientific time series. The approach addresses the gap between identifiability and interpretability in latent variable recovery. While existing CRL methods guarantee identifiability up to permutation and reparameterization, latent semantics are typically assigned post hoc. In scientific domains such as residue-pair distances, climate indices, or process sensors, observations are inherently semantic as they correspond to named physical quantities. MOSAIC integrates a sparse temporal VAE to transfer interpretability from observations to the latent space, enabling discovery of underlying causal structure without requiring ground-truth factors. The work is published on arXiv (2605.05524).