Abstract
<title>Abstract</title> <p>Seasonal forecasts support decision-making months ahead in agriculture, water, energy, and public health. Coupled dynamical systems like SEAS5 represent the state-of-the-art, but they are computationally expensive. While artificial intelligence (AI) offers a faster alternative, existing approaches have either predicted aggregated targets, prescribed past or future boundary conditions, or demonstrated freely evolving coupled dynamics without achieving broad multivariate seasonal forecast skill. We introduce the Functional Graph Transformer (FGT), a probabilistic AI forecasting system that freely evolves daily atmospheric and surface states including sea-surface temperature, sea ice, and soil moisture. In six-month forecasts, FGT outperforms a trend-aware climatology in 96% of evaluated metric-variable-lead combinations. In matched comparisons, it achieves at least 80% of SEAS5 skill in three quarters of cases while running several orders of magnitude faster. FGT's freely evolving surface state preserves useful El Niño–Southern Oscillation variability and boosts precipitation forecast skill, whereas persisting surface anomalies diminishes it. FGT also exhibits promising probabilistic behavior for selected high-impact events. The same model is competitive with specialist medium-range AI models at week 1 and outperforms a subseasonal AI model over weeks 1-6, suggesting that forecast skill across multiple horizons can emerge intrinsically, rather than requiring separate models.</p>