One diffusion model spans seventeen CMIP6 models: guidance trims the roughly threefold overestimate of extreme-heat coverage toward observations
Synopsis
Climate simulations are costly and biased, while correcting them too strongly toward observations risks erasing structural differences among models. A conditional diffusion model trained on seventeen CMIP6 models and ERA5 reanalysis steers each emulated model distribution toward observations at sampling time with an adjustable strength, chosen in perfect-model experiments with withheld simulated futures and decreasing with projection horizon and forcing. Across seventeen synthetic targets, guidance more than halves median biases in the 1st and 99th percentiles of daily temperature and brings CMIP6's roughly threefold historical overestimate of extreme-heat coverage close to observations.
Interpretation
A generative framework merges the simulated distributions of climate models with observed distributions to produce spatially coherent daily temperature fields, and uses them to estimate concurrent regional heat extremes. Climate simulations are computationally expensive and carry systematic biases relative to observations, yet correcting them directly toward observations risks suppressing inter-model differences and muting internal variability. In seventeen leave-one-model-out experiments using each held-out model's known future as a synthetic observation, guidance lowered 1st- and 99th-percentile daily temperature biases for all seventeen targets, averaged over three scenarios.
The merge is a weighted geometric mean of two learned distributions applied at sampling time, with the weight set by a guidance scale and no retraining. An additive mixture would draw each field from one distribution or the other, whereas the geometric mean keeps only fields plausible under both, which is what bias correction requires. A single conditional diffusion model represents the conditional distributions of the seventeen CMIP6 models and ERA5, and sampling integrates the probability-flow ordinary differential equation along the weighted sum of the two scores, the gradients of the log-densities.
Guidance strength is not fixed but declines linearly with calendar year and CO2-equivalent concentration, so farther futures rely increasingly on model dynamics. No future observations exist to tune the weight, so it is selected only in perfect-model experiments whose futures are known. The scale selected by minimizing 1st- and 99th-percentile biases falls from its 2025 value to lower values in 2100 under SSP2-4.5, SSP3-7.0 and SSP5-8.5 respectively.
Perspective
The demonstrated scope is daily near-surface temperature at moderate spatial resolution, covering seventeen CMIP6 models and ERA5 reanalysis over 1979–2100; the perfect-model experiments test out-of-model generalization only within the CMIP6 archive, so they cannot reveal biases shared across models or confirm calibration against observations. Independent daily sampling means multi-day heatwave duration, event sequencing and temporally compound risks cannot be estimated directly.
The guided ensembles remain somewhat under-dispersed: in the held-out MIROC6 fold the guided 5–95% warming interval covers the true trajectory in only part of the years, below the nominal coverage, and in the MPI-ESM1-2-LR target experiment the guided ensemble underestimates warming after mid-century. The authors note that tuning guidance strength is distinct from fully calibrating the predictive distribution, and that fully correcting remaining distributional errors would require test-time calibration algorithms.