A global empirical system for probabilistic seasonal climate prediction based on generative AI and CMIP6 models

Reliable probabilistic information at the seasonal time scale is essential across various societal sectors, such as agriculture, energy, or water management. Current applications of seasonal predictions rely on General Circulation Models (GCMs) that represent dynamical processes in the atmosphere, land surface, and ocean while capturing their linear and nonlinear interactions. However, GCMs come with an inherent high computational cost. In an operational setup, they are typically run once a month and at a lower temporal and spatial resolution than the ones needed for regional applications. Moreover, GCMs suffer from significant drifts and biases and can miss relevant teleconnections, resulting in low skill for particular regions or seasons. 

In this context, the use of generative AI methods that can model complex nonlinear relationships can be a viable alternative for producing probabilistic predictions with low computational demand. Such models have already demonstrated their effectiveness in different domains, i.e. computer vision, natural language processing, and weather prediction. However, although requiring less computational power, these techniques still rely on big datasets in order to be efficiently trained. Under this scenario, and with sufficiently high-quality global observational datasets spanning at most 70 years, the research trend has evolved into training these models using climate model output. 

In this work, we build upon the work presented by Pan et al., 2022, which introduced a conditional Variational Autoencoder (cVAE) to predict global temperature and precipitation fields for the October to March season starting from July initial conditions. We adopt several pre-processing changes to account for different biases and trends across the CMIP6 models. Additionally, we explore different architecture modifications to improve the model's performance and stability. We study the benefits of our model in predicting three-month anomalies on top of the climate change trend. Finally, we compare our results with a state-of-the-art GCM (SEAS5) and a simple empirical system based on the linear regression of classical seasonal indices based on Eden et al., 2015.

 

Pan, Baoxiang, Gemma J. Anderson, André Goncalves, Donald D. Lucas, Céline J.W. Bonfils, and Jiwoo Lee. 'Improving Seasonal Forecast Using Probabilistic Deep Learning'. Journal of Advances in Modeling Earth Systems 14, no. 3 (1 March 2022). https://doi.org/10.1029/2021MS002766.

Eden, J. M., G. J. van Oldenborgh, E. Hawkins, and E. B. Suckling. 'A Global Empirical System for Probabilistic Seasonal Climate Prediction'. Geoscientific Model Development 8, no. 12 (11 December 2015): 3947–73. https://doi.org/10.5194/gmd-8-3947-2015.