A Generative Deep Learning Model Complements Current Climate Model Emulators, Extending Monthly Outputs to Daily Frequency
A key scientific goal is to understand and project future climate changes under various human activity scenarios. Earth System Models (ESMs) are powerful tools for this purpose, but their computational demands limit the number of simulations that can be run. This constraint hinders the number of future scenarios that can be explored, and the robust analysis of risks associated with extreme weather events. To address this challenge, scientists are developing ESM emulators - computationally efficient models that can rapidly generate new realizations of climate projections once trained on an ESM-limited available output. However, many current emulators only provide output at monthly or coarser time scales, which is insufficient for analyzing daily extreme events like heat waves or heavy precipitation. To address this shortcoming, we have trained a novel ML algorithm that “downscales” the temporal frequency of the climate variables of interest from monthly to daily. We use diffusion modeling, which, though recently developed for video generation, has a versatility that enables a diverse set of applications for weather and climate modeling and emulation.
The DiffESM model presented in this paper provides a novel approach to emulating daily climate data from Earth System Models, adopting a popular machine learning approach originally developed for image generation. By learning the spatiotemporal characteristics of an ESM, using generative deep learning, DiffESM can effectively down-scale fields of monthly averages of temperature and precipitation to daily frequency. This allows for the rapid generation of daily temperature and precipitation data that closely matches the statistical characteristics of the ESM output, including the frequency and spatial patterns of extreme events. DiffESM enables researchers to quickly investigate the effects of various climate scenarios on daily-scale extreme weather events, complementing existing coarse-scale emulators and thus providing valuable insights into potential climate impacts and informing adaptation strategies.
DiffESM is a conditional diffusion model trained on a specific ESM’s output to generate daily temperature and precipitation data from monthly averages. Results show that DiffESM can produce daily climate data that closely matches the ESM output in terms of statistical distributions, spatial patterns, and extreme event characteristics. Importantly, DiffESM demonstrates strong performance across different forcing scenarios and time periods, including those not seen during training. When paired with an emulator that produces data on a monthly frequency, DiffESM allows for rapid generation of climate projections at a daily frequency, allowing for the investigation of extremes like heat waves, droughts, or heavy precipitation events.