Early-warning of the compact-to-dendritic transition via spatiotemporal learning from two-dimensional growth images

Transitions between distinct dynamical regimes are ubiquitous in nonequilibrium systems and are often preceded by weak, spatially heterogeneous precursors that are difficult to isolate from fluctuations. As a prototypical example, electrodeposition growth can undergo an irreversible compact-to-dendritic transition (CDT), marking the onset of morphological instability. Here, we formulate CDT prediction in a two-dimensional particle-based electrodeposition model as a horizon-based early-warning problem. We show that anticipating the transition is intrinsically spatiotemporal: static morphological descriptors, as well as temporal models built on predefined features, fail to provide reliable predictive signals over extended horizons. In contrast, end-to-end learning directly from growth-image sequences yields accurate anticipation across multiple horizons by jointly capturing evolving spatial structure and temporal dynamics. Analysis of the learned latent dynamics reveals an emergent low-dimensional surrogate coordinate that tracks progressive destabilization and reorganizes near the transition. We further show that the learned spatiotemporal representation exhibits limited but systematic transferability across reaction-rate conditions, with predictive performance degrading as the inference condition departs from the training condition, consistent with corresponding changes in the latent-state dynamics. These results establish a physics-grounded framework for forecasting incipient instabilities in pattern-forming nonequilibrium systems.