MERCURY: A fast and versatile multi-resolution based global emulator of compound climate hazards

TL;DR

MERCURY introduces a fast, multi-resolution climate hazard emulator that jointly represents regional monthly responses conditional on yearly GMT and reconstructs high-resolution grid-cell fields via a lifting-scheme operator. The framework combines a regression-based regional mean model with Monte Carlo sampling of wavelet patterns to preserve spatial correlations while enabling efficient zooming into regions of interest. Evaluation against CMIP6-based outputs shows MERCURY captures major spatial structures and ensemble spread for a WBGT proxy, though it underestimates some spatial correlations and exhibits regional quantile biases. The approach offers memory efficiency, scalability to multiple variables, and practical utility for rapid impact assessments under climate change, with clear paths for extension to more variables and non-Gaussian processes.

Abstract

High-impact climate damages are often driven by compounding climate conditions. For example, elevated heat stress conditions can arise from a combination of high humidity and temperature. To explore future changes in compounding hazards under a range of climate scenarios and with large ensembles, climate emulators can provide light-weight, data-driven complements to Earth System Models. Yet, only a few existing emulators can jointly emulate multiple climate variables. In this study, we present the Multi-resolution EmulatoR for CompoUnd climate Risk analYsis: MERCURY. MERCURY extends multi-resolution analysis to a spatio-temporal framework for versatile emulation of multiple variables. MERCURY leverages data-driven, image compression techniques to generate emulations in a memory-efficient manner. MERCURY consists of a regional component that represents the monthly, regional response of a given variable to yearly Global Mean Temperature (GMT) using a probabilistic regression based additive model, resolving regional cross-correlations. It then adapts a reverse lifting-scheme operator to jointly spatially disaggregate regional, monthly values to grid-cell level. We demonstrate MERCURY's capabilities on representing the humid-heat metric, Wet Bulb Globe Temperature, as derived from temperature and relative humidity emulations. The emulated WBGT spatial correlations correspond well to those of ESMs and the 95% and 97.5% quantiles of WBGT distributions are well captured, with an average of 5% deviation. MERCURY's setup allows for region-specific emulations from which one can efficiently "zoom" into the grid-cell level across multiple variables by means of the reverse lifting-scheme operator. This circumvents the traditional problem of having to emulate complete, global-fields of climate data and resulting storage requirements.

Figures (5)

• Figure 1: MERCURY's framework for generating monthly spatially multivariate climate fields. Yearly GMT values are used as inputs (panel a). The monthly, regional mean response and regional variability for each climate variable is first calculated (panel b). The lifting scheme is then employed to provide monthly spatially resolved, multivariate fields at the grid-cell level (panel c).
• Figure 2: Toy example of the lifting scheme applied on a grid consisting of 7 cells with values going from time 1 to $t$. At each iteration of the lifting scheme, the grid is split into groups of two (unless there is an odd number of cells in which case one group of three exists). The 'predict' step stores the wavelet coefficients representing the regression errors resulting from local regression (in our case, naive regression). Finally, the x values are updated with the scaling coefficients obtained by averaging values within each group. Split, predict and update steps are repeated until only a single scaling coefficient exists which corresponds to the grid average, and the wavelet coefficient grid is fully populated up to 6 cells (in the toy example's case two iterations). Within each lifting iteration the grid's spatial dimension is reduced by approximately a half.
• Figure 3: Difference between the ESM and emulator Spearman correlation matrix obtained by subtracting the emulator's Spearman correlation matrix from that of the ESM's for test scenario SSP 2-4.5. The Spearman correlation matrix is calculated over all land grid cells for January (upper triangle) and July (lower triangle).
• Figure 4: Comparison between the emulated and actual distribution for test scenario SSP 2-4.5. a) Quantile deviation maps for the 50$\%$, 95$\%$ and 97.5$\%$ quantiles, where red means that the emulated quantile is warmer than the actual ESM quantile and vice versa for blue. b) Probability rank distributions of the actual data with respect to the emulated ensemble. Data is aggregated to AR6 regional, continental and global levels before calculating the probability ranks. Whiskers indicate the 5th and 95th percentiles. If the distribution of actual data is captured perfectly, then the median should correspond to 0.5, edges to 0.75 and 0.25, and whiskers to 0.05 and 0.95.
• Figure 5: 2-D histogram time series for the emulated superensemble (greys), with the actual ESM superensemble overlaid for July, SSP 2-4.5. 1000 emulations were produced for each ESM initial-condition ensemble member. Note, that snce historical runs have more ensemble members, the emulation count is also higher up to 2015.