Latent Diffusion Model for Generating Ensembles of Climate Simulations

TL;DR

The paper tackles the high computational cost of generating large ensembles for climate uncertainty by introducing a latent diffusion framework that operates in a compressed latent space. It combines a pre-trained variational autoencoder (VAE) for dimensionality reduction with a denoising diffusion model (DDM) that learns latent residuals $$z_y = z - z_c$$ conditioned on a base latent $z_c = E(x_c)$, reconstructing samples as $$ hat{x} = D(z_c + hat{z}_y)$$ after diffusion-based generation. Two sequence-generation strategies are explored: an autoregressive method that incrementally predicts the next latent state and a transformer-based attention mechanism that processes the full time domain, enabling long-horizon climate simulations with controlled memory usage. On the MPI Grand Ensemble, the transformer-based approach closely matches the original ensemble's mean and variability and captures major climate features such as ENSO events and volcanic-induced shifts, while autoregressive generation provides strong temporal continuity. This approach offers a scalable, memory-efficient alternative for uncertainty quantification in climate projections and can be extended to additional variables, resolutions, and models to enhance decision-relevant climate risk assessments.

Abstract

Obtaining accurate estimates of uncertainty in climate scenarios often requires generating large ensembles of high-resolution climate simulations, a computationally expensive and memory intensive process. To address this challenge, we train a novel generative deep learning approach on extensive sets of climate simulations. The model consists of two components: a variational autoencoder for dimensionality reduction and a denoising diffusion probabilistic model that generates multiple ensemble members. We validate our model on the Max Planck Institute Grand Ensemble and show that it achieves good agreement with the original ensemble in terms of variability. By leveraging the latent space representation, our model can rapidly generate large ensembles on-the-fly with minimal memory requirements, which can significantly improve the efficiency of uncertainty quantification in climate simulations.

Figures (7)

• Figure 1: Monthly anomalies of the strongest El Niño events. Top row shows an original simulation of the MPI-GE, bottom row a selected generated simulation using our latent diffusion model.
• Figure 2: Our latent diffusion approach is split into two models, a variational autoencoder (VAE) pre-trained on independent climate states and a denoising diffusion model (DDM) trained on sequences of latent representations. During inference, the DDM generates new simulations in latent space, which are remapped to the original resolution by the decoder (D).
• Figure 3: Ensemble spread and ensemble mean of annual spatially averaged 98 original members from the MPI-GE (blue) compared to the generated members (red) from 1850 to 2005.
• Figure 4: ENSO timeline trenberth1997definition of an original simulation from the MPI-GE (blue) in comparison to a generated member (red) ranging from 1950 to 2005. The red dashed line marks the threshold of an El Niño event, the blue dashed line the threshold of a La Niña event.
• Figure 5: Ensemble spread and ensemble mean of annual spatially averaged 98 original members from the MPI-GE (blue) compared to the reconstructed members of the VAE (red) ranging from 1850 to 2005.