Hydro-JEPA: Towards Hydrological World Models Through Non-Generative Latent Embeddings

Abstract Hydrological modeling has traditionally relied on physically based models, conceptual rainfall-runoff formulations, regionalization techniques, and data-driven forecasting approaches. Recent advances in self-supervised learning and world models suggest an alternative paradigm: instead of explicitly defining hydrological behavior, artificial intelligence systems may learn latent representations of watershed dynamics directly from observations. Inspired by Joint Embedding Predictive Architectures (JEPA), this work proposes the concept of Hydrological Embeddings and a Hydrological World Model (Hydro-JEPA). Operating as a strictly non-generative architecture, Hydro-JEPA avoids the inefficient reconstruction of highly stochastic raw observations. Instead, it predicts the future purely in an abstract representation space, in which a watershed is represented by a latent state vector that evolves under climatic forcings (modeled as natural uncertainty) and human interventions (modeled as operational actions). We present a conceptual framework, mathematical formulation, and proof-of-concept methodology based on synthetic basins. The proposed approach aims not to replace traditional hydrology, but to provide a complementary representation capable of capturing hidden hydrological structures and dynamics. We hypothesize that latent hydrological states may constitute a new class of descriptors for watershed behavior, enabling transfer learning, scenario simulation, and integrated water resources planning.