ArchesClimate: Ensemble Generation for Decadal Prediction using Flow Matching

Sampling from climate models to generate ensembles of predictions is computationally expensive (Hawkins et al., 2015). Climate model ensembles are used to understand probabilities of climatic events and identify internal variability in climate models. In the short term, model uncertainty and inter-annual variability dominate uncertainty in climate predictions (Smith et al., 2019). A typical approach to address these uncertainties is to use large ensembles of non-learned, physical numerical global circulation models (GCM) (Eade et al., 2014). These ensembles allow for statistical analysis of distributions and determination of internal variability in the climate model.

Our approach demonstrates that we can efficiently learn to emulate a GCM. We use ensembles generated by the IPSL submission to the Decadal Climate Prediction Project (DCPP). The dataset ranges from 1960-2016 and produces 10-member, 10-year forecast ensembles for each year. On this dataset, we train a modified version of ArchesWeatherGen, a Swin Transformer based on PanguWeather that can be used in a generative way using flow matching (Couairon et al. 2024). The model was modified to predict additional climatic variables (e.g. air temperature, specific humidity, ocean potential temperature at depth, sea surface temperature, sea level pressure) at a monthly temporal resolution. Once trained, the model probabilistically generates ensemble members rapidly which can be auto-regressively rolled out. We show that they are physically reliable via evaluation methods that assess physical processes derived from the variables represented in the machine learning model, such as by evaluating it on El Niño/La Niña events. This model demonstrates that machine learning can enhance climate models by expanding ensemble sizes to improve our understanding of climatic processes. We aim to output physically realizable month-to-month trajectories to estimate future climate and its uncertainties across various domains, including land, ocean, and atmospheric processes.

Couairon, G., Singh, R., Charantonis, A., Lessig, C., & Monteleoni, C. (2024). ArchesWeather & ArchesWeatherGen: a deterministic and generative model for efficient ML weather forecasting. arXiv preprint arXiv:2412.12971.

Eade, Rosie, Doug Smith, Adam Scaife, Emily Wallace, Nick Dunstone, Leon Hermanson, et Niall Robinson. « Do Seasonal-to-Decadal Climate Predictions Underestimate the Predictability of the Real World? » Geophysical Research Letters 41, no 15 (2014): 5620‑28. https://doi.org/10.1002/2014GL061146.

Hawkins, Ed, Robin S. Smith, Jonathan M. Gregory, et David A. Stainforth. « Irreducible Uncertainty in Near-Term Climate Projections ». Climate Dynamics 46, no 11 (1 juin 2016): 3807‑19. https://doi.org/10.1007/s00382-015-2806-8.

Smith, D. M., R. Eade, A. A. Scaife, L.-P. Caron, G. Danabasoglu, T. M. DelSole, T. Delworth, et al. « Robust Skill of Decadal Climate Predictions ». Npj Climate and Atmospheric Science 2, no 1 (17 mai 2019): 1‑10. https://doi.org/10.1038/s41612-019-0071-y.