Diffusion-Based Joint Temperature and Precipitation Emulation of Earth System Models

TL;DR

This work extends a diffusion-based emulator (DiffESM) to jointly emulate temperature and precipitation, addressing the need for fast, statistically faithful climate realizations. The method conditions on monthly mean maps and outputs paired $28$-day sequences that preserve both intra-month coherence and the covariance between variables, enabling realistic assessment of extreme events. The approach is evaluated on IPSL-CM5A data under CMIP5 RCP8.5, showing close alignment with the original ESM in both marginal and joint distributions and coherent joint behavior across locations. The findings suggest substantial potential for multi-variable emulation to accelerate climate impact studies while retaining critical interdependencies, with plans to expand to additional variables and ESMs.

Abstract

Earth system models (ESMs) are the principal tools used in climate science to generate future climate projections under various atmospheric emissions scenarios on a global or regional scale. Generative deep learning approaches are suitable for emulating these tools due to their computational efficiency and ability, once trained, to generate realizations in a fraction of the time required by ESMs. We extend previous work that used a generative probabilistic diffusion model to emulate ESMs by targeting the joint emulation of multiple variables, temperature and precipitation, by a single diffusion model. Joint generation of multiple variables is critical to generate realistic samples of phenomena resulting from the interplay of multiple variables. The diffusion model emulator takes in the monthly mean-maps of temperature and precipitation and produces the daily values of each of these variables that exhibit statistical properties similar to those generated by ESMs. Our results show the outputs from our extended model closely resemble those from ESMs on various climate metrics including dry spells and hot streaks, and that the joint distribution of temperature and precipitation in our sample closely matches those of ESMs.

Figures (7)

• Figure 1: Training loop, illustrating the input and output channel $C=2$ for both daily temperature and precipitation.
• Figure 2: Joint variables discretized into temperature and precipitation deciles computed from the Held Out 2 realization. The distribution is computed only over wet days (when precipitation $>=1.00$ mm), for location Hawaii. The sets of contour maps contain no smoothing (left) and average smoothing with a $3 \hbox{by} 3$ kernel (right). Each set compares the distributions of the generated (left), Held Out 1 (middle), and Held Out 2 (right). Distributions are computed from 252 28-day samples.
• Figure 3: Joint distribution of variables discretized into temperature deciles computed from the Held Out 2 realization. The distribution is computed only over dry days (when precipitation $<1.00$ mm), for locations Hawaii (left), Melbourne (middle), Novosibirsk (right). The three bars of the figure compare the distributions of the generated realizations (yellow) to that of the Held Out 1 (green) and Held Out 2 (blue). Distributions are computed on the basis of 252 28-day samples.
• Figure 4: Pairs of difference maps between generated and Held Out 2 (left) and Held Out 1 and Held Out 2 realizations (right), and the corresponding superimposed grid-box error histograms of generated and Held Out 2 (orange) and Held Out 1 and Held Out 2 (blue) for two precipitation metrics (left) and two temperature metrics (right) for the bivariate sample conditioned on the Held Out 1 monthly means.
• Figure 5: Joint distribution of temperature and precipitation discretized into temperature and precipitation deciles computed from the Held Out 2 realization over wet days (when precipitation $>=1.00$ mm), for locations Melbourne (top), Novosibirsk (bottom). The set of contour maps on the left contain no smoothing, the set on the right contain average smoothing with a $3 \hbox{by} 3$ kernel. The three columns of the figure compare the distributions of the generated realizations (left) to that of Held Out 1 (middle) and Held Out 2 (right). Distributions are computed on the basis of 252 28-day samples.