Latent Diffusion and Spatiotemporal Transformers Generate Large Ensemble Climate Simulations
In one sentence. A VAE plus spatiotemporal-transformer latent diffusion model generates large climate ensembles on the fly from the MPI Grand Ensemble, avoiding storage of pre-computed simulations.
Abstract
Accurately estimating uncertainty in climate scenarios typically demands the generation of large ensembles of high-resolution simulations, a process that is both computationally expensive and memory intensive. To address this challenge, we present a novel deep learning framework that efficiently generates climate simulation ensembles by combining a variational autoencoder for dimensionality reduction with a denoising diffusion probabilistic model based on a spatio-temporal transformer architecture. This innovative approach not only captures the variability of the original ensemble, as validated on the Max Planck Institute Grand Ensemble, but also enables the rapid generation of large ensembles on-the-fly with minimal memory usage. This capability can transform climate research by eliminating the need to store vast amounts of pre-calculated simulation data, allowing for more extensive exploration of climate scenarios. These advances support more efficient uncertainty quantification, providing a broad range of possible outcomes that can effectively inform climate policy and adaptation strategies.
Labels
- Topic
- Climate emulation, Probabilistic & ensemble forecasting
- Type
- New model or method
- Method
- Diffusion / flow matching / score-based, Transformer / ViT, GAN or VAE
- Time scale
- Climate projections / centennial
- Domain
- Global
- Data
- CMIP / climate model output
- Status
- Research
- Identifiers
- DOI:10.1109/tai.2025.3621121 OpenAlex:W4415293143
- Added
- 2026-10-22
Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.