EarthOracle: A Self-Correcting World Model with Observational Feedback for Long-Horizon Earth Forecasting
Abstract
While deep learning models have achieved unprecedented success in short-to-medium-range weather forecasting, they universally suffer from error accumulation, spectral decay, and climate drift during long-horizon autoregressive rollouts. We argue that this stems from a fundamental paradigm flaw: treating forecasting as an open-loop ''one-shot extrapolation'' rather than the closed-loop ''predict-observe-assimilate'' cycle inherent in operational numerical systems. To address this, we propose EarthOracle, a self-correcting Earth World Model that explicitly embeds sparse observations into its inference loop. At its core lies a Differentiable Data Assimilation (Diff-DA) module, which dynamically injects heterogeneous and sparse observations (e.g., satellite radiances, buoy data, radiosondes) directly into the latent state using learnable uncertainty weights. Coupled with a novel self-correction training objective based on perturbation consistency, EarthOracle, actively digests new information to suppress error amplification over long trajectories. Furthermore, we introduce a Cross-Scale Memory mechanism that disentangles fast-evolving weather signals from slow-varying ocean-atmosphere coupled modes in the latent space, ensuring that observation feedbacks modulate the system at physically appropriate timescales. Extensive experiments on ERA5 (1979–2024) and multi-source sparse observations demonstrate that EarthOracle, significantly outperforms pure autoregressive baselines in Subseasonal-to-Decadal (S2D) forecasting. It maintains stable energy spectra and climatology over long rollouts, and matches the predictive skill of dynamical S2D models on critical teleconnection indices (e.g., ENSO, MJO) at less than 0.1% of their inference cost. By redefining long-term Earth forecasting from open-loop extrapolation to closed-loop assimilation, EarthOracle, provides a scalable and robust new paradigm for real-world AI climate prediction.