Skillful seasonal forecasts with a freely evolving AI weather model

Technical Summary: Skillful Seasonal Forecasts with a Freely Evolving AI Weather Model

Problem Statement
Seasonal forecasts are critical for decision-making in agriculture, energy, and public health but remain challenging due to the chaotic nature of atmospheric dynamics. Current state-of-the-art operational systems, such as the coupled dynamical model SEAS5, rely on computationally expensive atmosphere-ocean-sea-ice ensembles to generate probabilistic forecasts. While Artificial Intelligence (AI) has demonstrated success in medium-range weather prediction (up to two weeks) and subseasonal forecasting, existing AI approaches have failed to achieve broad, multivariate seasonal skill. Previous AI methods either predicted aggregated targets, prescribed boundary conditions (e.g., holding sea-surface temperature anomalies fixed), or demonstrated stability in long rollouts without quantifying broad forecast skill across the Earth System. There is a gap in achieving a single AI model that can freely evolve coupled atmospheric and surface states over six-month horizons while maintaining skill comparable to dynamical models.

Methodology: The Functional Graph Transformer (FGT)
The authors introduce the Functional Graph Transformer (FGT), a probabilistic AI forecasting system designed to evolve daily atmospheric and surface states autonomously for up to six months.

• Architecture: FGT utilizes a fully attention-based graph U-Net built upon the local icosahedral attention operator (Appa). It processes atmospheric, surface, radiation, sea-surface temperature (SST), sea-ice concentration, and soil moisture fields. The model advances daily states by taking the two preceding daily states as input. It employs self-attention within resolution levels and cross-attention between resolutions, with skip connections to propagate dynamics across scales. Unlike models that predict residual updates, FGT predicts the next state directly to avoid accumulating high-frequency artifacts during long rollouts.
• Probabilistic Generation: The system generates ensemble members using functional noise conditioning. Each member is conditioned on a distinct 64-dimensional noise vector sampled once and held fixed throughout the entire rollout to ensure temporal coherence.
• Training Strategy: To address the computational intractability of backpropagating through six months of autoregressive steps, FGT employs a "windowed pushforward" training scheme. The model is trained on 180-day autoregressive rollouts, but gradients are retained only within three specific scoring windows: near the start, a randomly sampled intermediate point, and near the end. This exposes the model to its own forecasts across weather, subseasonal, and seasonal lead times.
• Inference and Calibration: At inference, the system generates 300 candidate trajectories via a short pre-initialization window. These are ranked based on their ability to reproduce the observed initial conditions (using only data available at initialization). The top 51 candidates are retained to form the final ensemble, a process designed to improve ensemble calibration and reduce overdispersion.
• Data and Scope: The model is trained on ERA5 daily-mean fields (1959–2019) at 1° resolution and evaluated on the 2020–2026 period. It predicts 56 variables, including SST, sea ice, soil moisture, and various atmospheric pressure-level fields, without prescribing boundary conditions for the surface variables.

Key Contributions

1. First Broad-Skill Freely Evolving AI Model: FGT is presented as the first AI system to achieve broad seasonal skill (days to months) without prescribing boundary conditions, instead evolving the atmosphere, ocean surface, sea ice, and soil moisture autonomously.
2. Matched Comparison with SEAS5: The authors provide a rigorous, matched comparison against the operational SEAS5 model (same grid, initialization dates, ensemble size, and metrics).
3. Cross-Horizon Competence: The study demonstrates that a single model can maintain skill across weather (week 1), subseasonal (weeks 1–6), and seasonal (months 1–6) horizons, challenging the need for separate models for different timescales.

Results

• Skill Against Climatology: In six-month forecasts, FGT outperforms a trend-aware climatology in 96% of evaluated metric–variable–lead combinations (1,285 out of 1,344). This skill is distributed across atmospheric and surface fields, including precipitation, mean sea-level pressure, and winds.
• Skill Retention vs. SEAS5: In matched comparisons, FGT retains at least 80% of SEAS5's skill in 77% of cases and more than 50% in 92% of cases. While SEAS5 generally outperforms FGT in sea-surface temperature and near-surface temperature (particularly over oceans), FGT shows competitive or superior performance in precipitation and near-surface zonal wind over broad regions.
• Surface Evolution Impact: The study finds that allowing surface variables (SST, sea ice, soil moisture) to evolve freely preserves useful El Niño–Southern Oscillation (ENSO) variability and boosts precipitation skill. Conversely, persisting surface anomalies (a common practice in other AI seasonal studies) diminishes precipitation forecast skill.
• High-Impact Events: FGT demonstrates promising probabilistic behavior for specific high-impact events (e.g., the 2020 Indian heavy rainfall, 2022 East Africa drought), often capturing the sign and broad spatial pattern of anomalies better than SEAS5, though magnitude errors persist.
• Subseasonal Performance: FGT outperforms the subseasonal AI model FuXi-S2S over weeks 1–6 and remains competitive with medium-range specialists (GenCast, FGN) at week 1, suggesting that long-rollout training does not degrade shorter-term skill.
• Computational Efficiency: Generating a 51-member, six-month ensemble at 1° resolution takes approximately 19 minutes on a single NVIDIA A100 GPU. This represents a reduction of roughly six orders of magnitude in compute compared to extrapolating a coupled dynamical benchmark (IFS–NEMO).

Significance and Claims
The paper claims that FGT bridges the gap between weather, subseasonal, and seasonal forecasting, advancing the goal of "seamless prediction." The authors emphasize that FGT is not merely stable but skillful, achieving significant improvements over climatology and retaining substantial skill relative to state-of-the-art dynamical models.

The remaining gap between FGT and SEAS5 is presented as scientifically informative rather than a failure. The authors suggest that the degradation in SST and Dipole Mode Index skill indicates that surface-only ocean evolution is insufficient, pointing toward the need for subsurface ocean variables and richer land-state representations in future AI models. The study concludes that AI forecasts can sustain useful skill from days to seasons, opening a new frontier for long-horizon Earth-system prediction, provided that future work addresses the representation of ocean dynamics and improves long-range uncertainty calibration.