🤖 machine learning

Geospatial Diffusion-based Evolution Synthesis (GeoDES) for Storm-Centered Weather Augmentation

The paper introduces GeoDES, a geospatial diffusion-based image-to-video model that synthesizes high-fidelity, physically consistent storm structures to overcome the resolution and data limitations of existing weather models, achieving superior performance in predicting cyclonic storm dynamics.

Original authors: Sonia Cromp, Satya Sai Srinath Namburi GNVV, Youran Wang, Grace Kisslinger, Frederic Sala, James Booth, Allegra LeGrande

Published 2026-07-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sonia Cromp, Satya Sai Srinath Namburi GNVV, Youran Wang, Grace Kisslinger, Frederic Sala, James Booth, Allegra LeGrande

Original paper licensed under CC BY 4.0 (http://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

Imagine trying to predict the path of a giant, swirling whirlpool in the ocean, but you only have a camera that takes blurry photos of the entire ocean, or a camera that only looks at a tiny, fixed square of water. This is the current struggle for scientists trying to forecast extreme weather like massive cyclones. These storms are the "boss battles" of the atmosphere, responsible for the most dangerous winds and floods, yet they are notoriously hard to predict because they are chaotic and rare. To study them, scientists often use computer models, which are like digital simulations of the weather. However, these models face a tricky dilemma: if they try to simulate the whole world, they get too blurry to see the storm's sharp details; if they zoom in on just one storm, they often run out of computer power or get stuck because the storm moves out of their view. Furthermore, because extreme storms happen so rarely, computers don't have enough "practice" examples to learn how to predict them accurately. This is where a new type of artificial intelligence called a "diffusion model" comes in. Think of a diffusion model like a digital artist who learns to draw by starting with a messy pile of static (like the snow on an old TV) and slowly cleaning it up until a clear picture emerges. By teaching this artist to clean up weather data instead of TV static, scientists hope to generate new, realistic storm scenarios that can help them understand how these dangerous systems behave.

Enter GeoDES (Geospatial Diffusion-based Evolution Synthesis), a new AI tool designed by researchers to solve these specific problems. Instead of trying to predict the weather for the entire globe or a fixed city, GeoDES takes a "storm-centered" approach. Imagine you are filming a surfer riding a massive wave. Instead of filming the whole ocean and hoping the surfer stays in the frame, you put the camera on a drone that follows the surfer perfectly, keeping them right in the center of the shot. GeoDES does exactly this with cyclones: it locks onto the storm's low-pressure center and moves with it, treating the storm's life cycle like a video clip. This allows the model to focus all its computing power on the storm itself, ignoring the calm, boring weather around it.

The paper introduces a clever three-step training process to teach GeoDES how to make these videos. First, the AI learns to understand the "shape" of a storm in a single snapshot (like a still photo), mastering how wind and pressure relate to each other. Second, it uses a technique called "temporal inflation" to stretch that 2D knowledge into 3D, essentially teaching the AI how the storm moves and changes over time. Finally, it fine-tunes the whole system to generate a full video of the storm's evolution. The result is a model that can take a single starting image of a storm and generate the next 42 hours of its life, creating a new, realistic version of how that storm could have behaved.

The researchers tested GeoDES against several other advanced weather models, including some that try to predict the entire globe. The results were striking. While the global models tended to produce "blurry" storms that lacked sharp details (a problem known as spatial smoothing), and other models sometimes created chaotic, noisy nonsense, GeoDES managed to create storms that looked and felt physically real. In their tests, GeoDES achieved a 52% lower error in measuring the storm's spinning power (Peak Vorticity Error) and an 8% higher accuracy in matching the storm's pattern (Anomaly Correlation Coefficient) compared to the next best method. Perhaps most importantly, it was able to accurately recreate the intense energy of the wind, even for the most extreme storms, without inventing fake weather or smoothing out the dangerous parts.

The paper also highlights that this high level of detail didn't require a supercomputer. Because GeoDES only focuses on the storm's immediate area, it can run on standard consumer-grade graphics cards, making it much more accessible than the massive global models that need huge data centers. The authors show that by using their specific training steps, they could build a model that is both highly accurate and computationally efficient. They explicitly rule out the idea that simply making a bigger global model is the answer, showing that those models often fail to capture the fine-grained details of cyclones. Instead, they demonstrate that focusing strictly on the storm's evolution, while keeping the physics consistent, is the key to generating realistic, high-fidelity weather data. This work suggests that we can now create better "practice storms" to test our forecasting tools and expand our understanding of extreme weather, all without needing to simulate the entire planet.

Technical Summary: Geospatial Diffusion-based Evolution Synthesis (GeoDES)

Problem Statement

Machine learning (ML) based weather models face significant structural challenges when predicting the detailed dynamics of large-scale extratropical cyclones. Current architectures are generally constrained to either global or regional prediction scopes, both of which are ill-suited for tracking detailed, far-moving storm systems.

Global Models: While comprehensive, they are computationally expensive and often operate at resolutions too coarse to capture fine-grained storm dynamics. To remain tractable, they frequently produce heavily smoothed predictions or rely on masking strategies that introduce spurious dynamics.
Regional Models: These are restricted to static geographic boundaries, making them unable to track storms that move across large distances. Furthermore, both global and regional approaches waste significant training capacity on modeling calm, non-stormy conditions.
Data Scarcity: Extreme storms are inherently difficult to study due to limited observational sample sizes, causing ML methods to underperform on rare, high-impact events compared to traditional physics-based forecasting.

Existing ML weather models struggle to preserve the rotational structure and kinetic energy distributions of these storms, often succumbing to spatial smoothing or generating unphysical "hallucinations."

Methodology: GeoDES

To address these limitations, the authors introduce GeoDES (Geospatial Diffusion-based Evolution Synthesis), a custom image-to-video diffusion model designed specifically for storm-centered weather augmentation. The core innovation lies in a "storm-centered" approach where each cyclone is treated as a video anchored to its moving low-pressure center, rather than a static grid.

Architectural Pipeline

GeoDES is constructed via a three-phase pipeline to decouple spatial structure from temporal evolution, preventing the physical hallucinations and dynamic range collapse seen in models trained from scratch:

Phase 1: Two-Dimensional Spatial Pretraining:
- A 2D U-Net ( $f_{\Theta2}$ ) is trained on flattened spatiotemporal data to learn a Spatial Prior Network.
- This phase focuses on capturing multivariate atmospheric relationships (e.g., pressure gradients and wind fields) without the complexity of temporal dynamics.
- Hybrid Normalization: To prevent dynamic range collapse, the model uses log-scaling for long-tailed distributions (e.g., wind magnitude) and robust linear scaling for centered distributions (e.g., temperature), scaled by the 1st and 99th percentiles.
- SNR Weighting: The training objective uses Signal-to-Noise Ratio (SNR) weighting to force the model to dedicate capacity to meaningful, large-scale storm structures rather than high-frequency noise.
Phase 2: Temporal Inflation:
- The 2D spatial weights are transferred to a 3D Spatiotemporal Evolution Network ( $f_{\Theta3}$ ) using a "temporal inflation" strategy.
- Spatial blocks are copied directly, while weights requiring a temporal dimension are initialized by repeating the 2D weights across time steps ( $w_3(t) = w_2/T$ ). This preserves the learned atmospheric representations while introducing the time dimension.
Phase 3: Three-Dimensional Spatiotemporal Fine-Tuning:
- The inflated 3D model is fine-tuned on temporal dynamics.
- Correlated Noise: Unlike standard diffusion which uses independent Gaussian noise per frame, GeoDES injects correlated noise ( $\epsilon(t) = \rho\epsilon(t-1) + \sqrt{1-\rho^2}\epsilon_u(t)$ ) to respect the high autocorrelation of atmospheric states and prevent physically impossible state jumps.
- Non-Autoregressive Sampling: The model predicts the entire $T-1$ future timesteps simultaneously conditioned on an initial observation (prompt). During sampling, the un-noised initial frame is re-injected at every step to anchor the generation to the prompt.

Key Contributions

Task Formalization: The authors formalize the data augmentation task of conditional storm synthesis, focusing on generating physically consistent storm evolutions from initial conditions.
Novel Architecture: GeoDES introduces a spatiotemporal diffusion model that utilizes storm-centered framing and temporal inflation. This bypasses the computational limits of global models and the spatial smoothing of regional models.
State-of-the-Art Generative Realism: The model demonstrates superior performance in preserving rotational structure and extreme wind energy distributions compared to existing baselines.
Extreme Event Augmentation: GeoDES successfully generates physically consistent severe storms, enriching the training distribution for rare, high-intensity events without succumbing to the smoothing or hallucinations common in global models.

Results and Evaluation

Evaluated on North Atlantic extratropical cyclones (2016–2024), GeoDES outperforms a diverse suite of baselines, including statistical climatology, off-the-shelf diffusion models (Stable Video Diffusion), and pretrained Weather Foundation Models (WFMs) like ClimaX and Aurora.

Performance Metrics

Physical Fidelity: GeoDES achieved a 52% lower Peak Vorticity Error and an 8% higher Anomaly Correlation Coefficient (ACC) than the next strongest methods on the North Atlantic test set.
Energy Distribution: It is the only model capable of accurately recovering the kinetic energy distribution of ground-truth storms, achieving a near-perfect Frequency Bias Index (FBI) of 1.02 and a High-Frequency (HF) Spectral Ratio of 0.95.
Extreme Storms: On the top 10% most extreme storms, GeoDES maintained its performance, correctly preserving the rotational core (Peak Vorticity Error of 4.47) and recovering extreme wind distributions. In contrast, global WFMs exhibited catastrophic spatial smoothing (FBI of 0.00), while off-the-shelf diffusion models failed to constrain physics, generating excessive noise.

Computational Efficiency

Accessibility: Despite having a large parameter count (1.8B), GeoDES's localized strategy restricts computation to the storm's bounding box. This allows training and inference on standard consumer-grade GPUs.
Comparison: While global WFMs like Aurora have fewer parameters (1.25B), their requirement to process high-resolution global grids results in over 36 times more FLOPs and nearly double the VRAM usage compared to GeoDES.
Training Efficiency: The temporal inflation strategy reduces the total theoretical compute required for training by approximately 46% compared to training a 3D architecture from scratch.

Significance

The paper claims that GeoDES provides a critical bridge between the computational constraints of global models and the resolution needs of detailed storm analysis. By focusing strictly on the evolving storm structure, GeoDES enables the synthesis of realistic, high-fidelity weather events suitable for stress-testing forecast models and expanding meteorological datasets.

The authors emphasize that this approach is particularly vital for studying extreme events, which are underrepresented in historical data. By generating physically consistent severe storms, GeoDES offers a tool to enrich the training distribution for downstream climate studies, addressing the "double penalty" problem where pixel-wise metrics unfairly penalize realistic but slightly displaced features. The work suggests that storm-centered, non-autoregressive diffusion models represent a viable path forward for high-resolution, computationally accessible weather synthesis.