CloudFlow: A Flow Matching Model to Generate High-Resolution Cloud Structures

Tim Reichelt, Philip StierAI4Science @ ICML 2026 · OpenReview

TL;DR

CloudFlow is a new generative model that uses conditional flow matching to produce high-resolution (~1km) cloud structures from coarse-scale (~25km) atmospheric data. By learning to "hallucinate" realistic clouds based on 3D atmospheric profiles, it provides a computationally efficient way to study cloud feedbacks—the single largest source of uncertainty in our climate change projections.

What Problem Does This Paper Address?

Predicting how much the Earth will warm (Equilibrium Climate Sensitivity) is a high-stakes game of "what if." Currently, our estimates range from a manageable 2 Kelvin to a catastrophic 5 Kelvin. The primary reason for this massive uncertainty? Clouds.

Global Climate Models (GCMs) typically operate at resolutions of 25–100 kilometers. However, the processes that govern cloud formation and organization happen at the "mesoscale" (5–100 km). Because GCMs can’t "see" these small-scale structures, they rely on simplified approximations. While high-resolution "storm-resolving" models exist, they are so computationally expensive that running them for century-scale climate predictions is currently impossible.

Researchers often use "Cloud Controlling Factors" (CCFs)—hand-designed physics features—to bridge this gap, but these human-defined shortcuts often discard vital information hidden in the full atmospheric profile.

Methods

The authors introduce CloudFlow, a generative model based on the Flow Matching formalism. Unlike standard diffusion models, flow matching learns to map a simple probability distribution (noise) to a complex one (satellite images) by learning the vector field that connects them.

Key Innovations:

  • Attention-Based Preprocessing: The model doesn't just look at surface temperature. It takes in full 3D atmospheric profiles (temperature, humidity, wind) from ERA5 reanalysis data. The authors developed an Attention Transform that groups variables by pressure level, allowing the model to learn which parts of the atmospheric column are most relevant for specific cloud types.
  • Conditional Generation: The model is "conditioned" on coarse-scale data. It doesn't just generate any cloud; it generates the specific cloud structures that are physically consistent with the surrounding large-scale weather patterns.
  • U-Net Backbone: It utilizes a U-Net architecture (borrowed from the CorrDiff model) to handle the spatial downscaling from 25km atmospheric grid cells to 1km cloud pixels.

Datasets

The model was trained and evaluated on two primary sources:

  1. Conditioning Data: ERA5 reanalysis data, providing atmospheric profiles at 37 different pressure levels.
  2. Target Data: 20 years of MODIS Aqua satellite observations. This includes both raw calibrated radiances (visible and infrared) and "Level 2" retrieved properties like Cloud Top Height (CTH), Cloud Water Path (CWP), and Cloud Optical Thickness (COT).

The training set consisted of 40,000 global tiles (128x128 pixels) from 2003–2011, with 1,000 tiles from 2018 reserved for testing.

Key Insights and Ideas

  • Learned vs. Hand-Crafted Physics: The "Attention" version of CloudFlow consistently outperformed a baseline model that used traditional, hand-designed CCFs. This suggests that deep learning can discover "latent" cloud-controlling factors that humans haven't explicitly defined yet.
  • Spectral Fidelity: One of the biggest wins for generative models in this field is their ability to produce "sharp" images. While traditional regression models often produce blurry, averaged-out results, CloudFlow’s generations match the power spectra of real satellite data even at high frequencies (small scales).
  • Physical Consistency: The model demonstrates "cross-channel" logic. For instance, when it generates a cloud with a high Cloud Top Height, it correctly generates a correspondingly cooler brightness temperature in the infrared bands—a sign that it is learning the underlying physics, not just memorizing textures.

Key Results

  • Accuracy: The Attention-based model achieved significantly lower Mean Absolute Error (MAE) and better Continuous Ranked Probability Scores (CRPS) than both the Linear model and the CCF baseline across almost all variables.
  • Cloud Top Height: The model's error for Cloud Top Height is roughly 2 km, which is remarkably on the same order of magnitude as the inherent retrieval errors of the MODIS satellite itself.
  • Speed: While traditional Large Eddy Simulations (LES) can take hours on a supercomputer to simulate a small domain, CloudFlow can generate samples in seconds on a single NVIDIA A100 GPU.

Limitations and Future Work

  • Underdispersive Predictions: The model is slightly "underdispersive," meaning it doesn't always capture the full range of possible cloud outcomes for a given atmospheric state (it's a bit too "confident" in its average guess).
  • The "Feedback" Missing Link: While the model generates clouds, it doesn't yet directly predict their radiative effects (how much sunlight they reflect vs. heat they trap). The authors aim to extend the model to predict these effects directly.
  • Interpretability: The next step is to use "mechanistic interpretability" tools to peek inside the neural network and see if the features it learned match known meteorological concepts or if it has discovered entirely new precursors to cloud formation.
  • Temporal Dynamics: Currently, the model uses MODIS data, which only provides a snapshot at a specific time of day. Future versions could incorporate geostationary satellite data to capture the full diurnal cycle of clouds.

Summary updated

Abstract

Our limited understanding of clouds is the dominant source of uncertainty in future climate predictions. Understanding how changing atmospheric environmental conditions constrain cloud organisational patterns and their radiative effects is key to understanding their impact on future climate change. We present CloudFlow, a flow matching model that is able to generate high-resolution cloud structures conditioned on coarse-scale atmospheric conditions. Our model generates realistic cloud structures that match the spectra and distributions of the original high-resolution scenes. CloudFlow introduces a new modeling regime to study how atmospheric environmental conditions impact cloud morphologies which will contribute to an improved understanding of cloud feedbacks, the cloud response to a changing climate and its effect on climate itself.
BibTeX
@inproceedings{reichelt2026cloudflow,
  author = {Tim Reichelt and Philip Stier},
  title = {{CloudFlow: A Flow Matching Model to Generate High-Resolution Cloud Structures}},
  booktitle = {AI4Science @ ICML},
  year = {2026},
  url = {https://openreview.net/forum?id=meHQFrgQqw},
}
Figure