Toward Generative Machine Learning for Boosting Ensembles of Climate Simulations
Parsa Gooya, Reinel Sospedra‐Alfonso, Johannes Exenberger
Abstract Accurately quantifying uncertainty in predictions and projections arising from irreducible internal climate variability is critical for decision‐making. Such uncertainty is typically assessed using ensembles produced with climate models. However, computational constraints impose a trade‐off between generating large ensembles required for robust uncertainty estimation and increasing model resolution to better capture fine‐scale dynamics. Generative machine learning offers a promising pathway to alleviate these constraints. We develop a conditional Variational Autoencoder (cVAE) trained on a limited sample of climate simulations to generate arbitrary large ensembles. The approach is applied to output from monthly CMIP6 historical and scenario experiments produced with the Canadian Centre for Climate Modelling and Analysis' Earth system model CanESM5. We show that the cVAE model learns the underlying distribution of data and generates physically consistent samples that reproduce realistic low‐ and high‐moment statistics, including extremes. Compared with more sophisticated generative architectures, cVAEs offer mathematically transparent, interpretable, and computationally efficient framework. Their simplicity lead to some limitations, such as smooth outputs, spectral bias, and underdispersion, that we discuss along with mitigation strategies. Specifically, we show that incorporating output noise improves the representation of climate‐relevant multiscale variability, and propose a simple method to achieve this. We show that cVAE‐enhanced ensembles capture realistic global teleconnection patterns, even under climate conditions absent from training data. Finally, our results point to limitations in accurately capturing non‐Gaussianity in higher‐order moments. It remains to be addressed whether this is amendable via more expressive architectures and output noise treatment or remains a challenge with cVAEs more generally.