📄 earth science

Towards Forecasts of Seasonal Tropical Cyclone Activity in the North Atlantic using Diffusion Models

This paper presents a conditional multivariate diffusion model that effectively downscales low-resolution monthly seasonal forecasts into realistic, high-resolution weather snapshots, demonstrating significant potential for improving tropical cyclone risk assessment in the North Atlantic despite some limitations in capturing specific warm-core systems.

Original authors: Benjamin Ulmer, Charles Powell, Robert Rouse

Published 2026-10-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Benjamin Ulmer, Charles Powell, Robert Rouse

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Hurricanes are among the most powerful and destructive forces on our planet, capable of reshaping coastlines and causing billions of dollars in damage in a single season. For communities living in their path, the difference between safety and disaster often comes down to anticipation. Meteorologists have long been able to look months ahead and predict the broad conditions that favor storm activity, such as warm ocean waters or specific shifts in global wind patterns. These seasonal outlooks are invaluable for preparing infrastructure and managing risk, but they come with a significant blind spot. Because they average weather over entire months, these forecasts smooth out the very details that matter most: the swirling winds, the pockets of intense moisture, and the specific structures that allow a tropical disturbance to grow into a hurricane. They tell us the stage is set for a storm, but they cannot show us the storm itself.

A team of researchers at the University of Cambridge has developed a new method to bridge this gap, turning those broad monthly predictions into detailed, realistic snapshots of daily weather. Their work focuses on the North Atlantic, a region where hurricanes frequently form and travel. The scientists used a type of advanced computer model known as a diffusion model. In simple terms, this technology learns to fill in the missing details of a blurry picture. The researchers trained the model using decades of historical weather data, teaching it the relationship between the smooth, low-resolution monthly forecasts and the complex, high-resolution weather patterns that actually occur on the ground. The goal was not to predict a specific storm months in advance, which remains impossible, but to generate a collection of plausible weather scenarios that are consistent with a given seasonal forecast. By creating many different possible outcomes, the model provides a much richer picture of what the season might look like, allowing for a more accurate assessment of the risks posed by tropical cyclones.

The researchers tested their system by feeding it monthly forecasts from the European Centre for Medium-Range Weather Forecasts, a leading global weather organization. For five hundred different dates between 2022 and 2024, the model generated twenty different six-hourly weather snapshots for each date. These snapshots included twelve different atmospheric variables, such as wind speed, temperature, humidity, and air pressure, all mapped across the North Atlantic. The results were striking. The generated weather fields captured the fine-scale details that the original monthly forecasts lacked, restoring the natural variability of the atmosphere. The model successfully recreated the complex relationships between different weather elements, such as how wind patterns and air pressure interact, without being explicitly programmed with the laws of physics. Instead, it learned these connections directly from the historical data, producing fields that were statistically and physically consistent.

When the researchers examined these generated snapshots for signs of tropical cyclones, the model performed with impressive accuracy. It identified storm-like structures at a frequency very close to what was observed in the historical record, capturing the seasonal rhythm of hurricane activity with a high degree of correlation. The model correctly reproduced the peak of the hurricane season in late summer and generated storms that possessed the key physical characteristics of real hurricanes, including the warm cores and rotating winds that define them. However, the system was not perfect. It tended to produce slightly fewer storms than the historical record when applying a strict definition of a hurricane, and it showed a slight bias toward creating storms in the cooler, northern parts of the ocean rather than the warm tropical waters where they typically form. Additionally, the collection of generated storms was slightly less varied than reality, meaning the model sometimes produced similar-looking outcomes rather than a wide range of possibilities.

Despite these minor limitations, the study demonstrates a significant step forward in how we can use seasonal forecasts. By translating coarse, long-range predictions into detailed, high-resolution weather fields, this approach offers a new tool for understanding the potential severity of an upcoming hurricane season. It moves beyond simply saying that a season will be active to showing what that activity might actually look like. This capability is crucial for risk assessment, as it allows communities and asset owners to visualize the specific weather conditions they might face. While the model does not predict individual storms months in advance, it provides a realistic ensemble of possibilities that can inform better preparation and response strategies. The work suggests that as these models improve, they could eventually be used to explore how climate change might alter the frequency and intensity of hurricanes, offering a clearer view of the future risks facing coastal regions.

Technical Summary: Towards Forecasts of Seasonal Tropical Cyclone Activity in the North Atlantic using Diffusion Models

Problem Statement
Anticipating tropical cyclone (TC) activity is critical for risk management, yet current seasonal forecasting systems face a resolution gap. While systems like the European Centre for Medium-Range Weather Forecasts (ECMWF) SEAS5 provide valuable long-range (up to 7 months) predictions of large-scale drivers (e.g., SST, ENSO), they output monthly mean fields. These coarse, time-averaged forecasts lack the high-resolution spatial variability and instantaneous storm structures necessary to assess TC risk. Existing downscaling methods often rely on hard-coded physical assumptions or stochastic seeding that may not fully capture the complex, nonlinear interactions governing storm genesis. There is a need for a method that can translate low-resolution seasonal forecasts into realistic, high-resolution, multivariate weather snapshots capable of supporting TC detection algorithms.

Methodology
The authors propose a conditional multivariate diffusion model designed to temporally downscale monthly SEAS5 forecasts into six-hourly weather snapshots over the North Atlantic.

Data and Domain: The study focuses on the North Atlantic (0°–59.75°N, 85°–9.25°W). The model is trained on pairs of interpolated SEAS5 monthly mean forecasts (conditioning variables) and corresponding ERA5 reanalysis data (targets). The dataset spans 1981–2018 for training, with 2019–2021 for validation and 2022–2024 for testing.
Residual Learning: Instead of generating full weather fields directly, the model learns to predict the residual between the ERA5 weather and the interpolated SEAS5 baseline. This approach reduces the variance the model must learn, allowing it to focus on resolving fine-scale structures missing from the monthly means.
Model Architecture: A U-Net with five levels and approximately 160 million parameters is used. The network predicts the velocity of the diffusion process ( $v$ -prediction) rather than the noise, which improves stability at high noise levels.
Conditioning and Inputs: The reverse diffusion process is conditioned on the normalized SEAS5 baseline and a land/sea mask. The model generates twelve variables jointly: horizontal winds ( $u, v$ ), temperature, relative humidity, and relative vorticity at 850 and 500 hPa, plus Mean Sea Level Pressure (MSLP) and Sea Surface Temperature (SST).
Training Enhancements:
- Offset Noise: Added to the noise schedule to ensure the model can reproduce non-zero mean residuals, which is crucial for climate downscaling.
- Temporal Smearing: During training, ERA5 timestamps are displaced by up to 18 hours relative to the SEAS5 baseline to expose the model to short-term variability under the same seasonal conditions.
- Loss Function: A combination of L1 loss (robust to extreme events) and a distribution penalty (Kullback–Leibler divergence) to capture long tails in the data, weighted by Min-SNR to prevent low-noise timesteps from dominating training.
Sampling: The model uses Denoising Diffusion Implicit Models (DDIM) with 500 steps and $\eta=0.5$ to generate an ensemble of 20 independent samples per date.
TC Detection: Generated samples are evaluated using the Okubo–Weiss–Zeta (OWZ) algorithm to identify TC-like candidates. An additional thickness criterion (approximating geopotential thickness between 850 and 500 hPa) is applied to filter for warm-core systems.

Key Results
Evaluated on 500 dates from 2022–2024, the diffusion model demonstrates significant improvements over the interpolated SEAS5 baseline:

Distributional Accuracy: The generated samples reproduce ERA5 marginal distributions more closely than the baseline for 11 of 12 variables. The mean normalized Wasserstein distance improved from 0.3140 (baseline) to 0.0588 (generated).
Spatial Variability: The model successfully restores spatial variability absent in the monthly means. The log-spectral distance (LSD) improved significantly, particularly in the Meso- $\alpha$ band (200–2000 km), where the baseline is weakest. The generated fields capture storm-scale circulations.
Physical Consistency: Despite not being explicitly enforced in the loss function, the model learned the physical relationship between horizontal winds and relative vorticity, with regression slopes near 1.0 and $R^2 > 0.99$ .
Probabilistic Calibration: The ensemble is slightly under-dispersed (mean spread-skill ratio of 0.871), indicating a tendency to produce similar samples and potentially underestimate uncertainty. However, the Continuous Ranked Probability Score (CRPS) improved for all variables compared to the deterministic baseline.
TC Candidate Recovery:
- The model recovers 94.8% of the total OWZ candidate detections found in ERA5.
- Monthly variation in candidate counts correlates highly with ERA5 ( $r=0.969$ ), capturing the seasonal peak in late summer.
- When applying the stricter warm-core (thickness) criterion, the model underrepresents candidates (37.8 detections vs. 121 in ERA5) and shows a bias toward extratropical systems, particularly in the northwestern and southeastern Atlantic.
- Structural analysis of TC-like features shows underestimation of central pressure deficits and maximum wind speeds, with a slightly larger radius of maximum winds compared to ERA5.

Significance and Claims
The paper claims to demonstrate the potential of diffusion models to translate seasonal forecasts into realistic, multivariate weather fields suitable for tropical cyclone risk assessment. The authors emphasize that the objective is not to predict specific individual storms months in advance, but to generate an ensemble of plausible weather conditions consistent with predicted seasonal drivers.

Key contributions include:

Proof of Concept: Showing that diffusion models can effectively downscale coarse seasonal forecasts to high-resolution, instantaneous weather snapshots that preserve physical relationships between variables.
Statistical Fidelity: Establishing that the generated ensembles recover the frequency, seasonality, and spatial distribution of TC-like structures found in reanalysis data, despite the inherent limitations of independent sampling.
Methodological Advancement: Utilizing residual learning and offset noise to address specific challenges in climate downscaling, such as reproducing mean states and capturing extreme event tails.

The authors acknowledge limitations, noting that the model generates independent snapshots (preventing track reconstruction), relies on ERA5 (a reanalysis) as ground truth, and currently exhibits a bias toward extratropical systems. They suggest future work could involve conditioning on climate change scenarios (CMIP6) or extending the model to generate temporally coherent sequences via auto-regressive conditioning.

More like this