Statistical Downscaling and Emulators: Can Generative Machine Learning add Value to Extrapolation and Bias?

Besides regional climate models (RCMs), there exist two main approaches to tackle the insufficient resolution of global climate models: emulators and statistical downscaling. While both approaches are similar in the techniques they use (statistical and machine learning, ML, methods), they differ in their objectives and underlying assumptions. Emulators are intended to provide a cost-effective alternative to RCMs by emulating their downscaling functions. Alternatively, statistical downscaling (SD) models learn the empirical (observed) relationships that link a set of key large-scale predictors, to the local high-resolution predictand of interest. There is a key tradeoff between these two approaches: emulators are unconstrained by observed climate records, yet they also inherit RCM biases; conversely, SD methods are able to produce potentially bias-free simulations (at least when driven by reanalyses), but with extrapolation constrained by observed relationships.

This tradeoff between extrapolation and bias is a key research perspective, especially when compounded with the usual additional challenges ML methods face, like representation of extremes or the temporal/spatial consistency of the predictions. Within this context, the added value of generative/stochastic methods is highly relevant and timely. Indeed, recent studies using deterministic ML methods (such as Wang et al. 2023; Doury et al. 2024) have highlighted that emulating high-resolution fields does require generative/stochastic approaches, specially when it comes to representing extreme weather events for complex variables like precipitation (Watson, 2022, 2023). However, while generative methods such as diffusion models may offer an advantage when it comes to simulating extremes (Addison et al. 2024; Aich et al. 2024), they are also subject to more potential instability (e.g., diffusion models are known to have hallucinations, Aithal et al. 2024), hence also increasing the biases.

In this study we aim to address the added value and potential downsides generative/stochastic ML methods can bring to the field of statistical downscaling and emulation, by targeting the tradeoff between extrapolation and bias. Therefore, we will address both already well-established generative deep learning techniques and the latest generation diffusion models, and focus on how well they fare when capturing aspects beyond mean statistics, including extremes, which are of particular interest in terms of climate impacts.

 

References:
Addison, H. et al. (2024). Machine learning emulation of precipitation from km-scale regional climate simulations using a diffusion model. Preprint. DOI: https://doi.org/10.48550/arXiv.2407.14158

Aich, M. et al. (2024). Conditional diffusion models for downscaling & bias correction of Earth system model precipitation. Preprint. DOI: https://doi.org/10.48550/arXiv.2404.14416

Aithal, S. K. et al. (2024). Understanding Hallucinations in Diffusion Models through Mode Interpolation. Preprint. DOI: https://doi.org/10.48550/arXiv.2406.09358

Doury, A. et al. (2024). On the suitability of a convolutional neural network based RCM-emulator for fine spatio-temporal precipitation. Climate Dynamics, 62(9), 8587-8613. DOI: https://doi.org/10.1007/s00382-024-07350-8

Watson P. A. G. (2022). Machine learning applications for weather and climate need greater focus on extremes. Environmental Research Letters 17(11). DOI: https://doi.org/10.1088/1748-9326/ac9d4e

Watson, P. (2023). Machine learning applications for weather and climate predictions need greater focus on extremes: 2023 update. NeurIPS 2023 Workshop on Tackling Climate Change with Machine Learning.

 

Acknowledgement:

This work is funded by the National Research Agency under France 2030 bearing the references ANR-22-EXTR-0005 (TRACCS-PC4-EXTENDING project) and ANR-22-EXTR-0011 (TRACCS-PC10-LOCALISING project).