Probabilistic Global Ocean Forecasting Through Diffusion-Based Ensembles

Accurate ocean forecasting is essential for a range of critical applications, from maritime safety to climate adaptation strategies. Given the inherent uncertainties in ocean dynamics, the ability to predict a range of probable ocean states is key to informed decision-making. Here, we present MerCast, a probabilistic ocean forecasting model designed to redefine global-scale prediction by quantifying uncertainty in ocean state estimates. Trained on decades of high-resolution reanalysis products, MerCast integrates diffusion models to generate ensembles of daily forecasts at 1/4-degree resolution, dynamically capturing local-global interactions while preserving fine-scale ocean features essential for accurate predictions.

MerCast's  performance is rigorously evaluated using an array of metrics tailored for stochastic forecasting systems, including ensemble spread, probabilistic error assessments, and metrics designed for process-oriented evaluations. Initial results highlight MerCast's skill in forecasting critical variables such as sea surface height, temperature, salinity, and ocean currents, with superior resilience to error accumulation over extended forecast horizons. This work establishes a foundational step toward integrating probabilistic methods in operational ocean forecasting, bridging the gap between efficiency, accuracy, and uncertainty quantification.