Efficient Large Ensemble Generation of Climate Model Output Using Latent Diffusion and Spatio-Temporal Transformers

Estimating uncertainty in climate scenarios often requires generating large ensembles of high-resolution simulations, a task that is both computationally and memory intensive. To overcome these challenges, we propose a deep learning framework that combines a variational autoencoder for dimensionality reduction with a denoising diffusion probabilistic model built on a spatio-temporal transformer architecture. The model is trained on large ensembles of low-resolution climate model outputs to capture internal variability and a single high-resolution climate model output to generate high-resolution simulations. This innovative approach enables the dynamic generation of large ensembles of high-resolution simulations with minimal computational overhead, eliminating the need for storing extensive precomputed data. By facilitating the efficient quantification of uncertainty, this framework provides a powerful tool for exploring a wide range of high-resolution climate outcomes, supporting the development of informed climate policies and adaptation strategies.