Guiding future global temperature distributions with simulations and observations

Francesco Immorlano, Elijah Tavares, Felix Draxler, Juan Nathaniel, Violette Launeau, Duncan Watson-Parris, Padhraic Smyth, Pierre Gentine, Stephan MandtarXiv 2026 · arXiv

TL;DR

Researchers have developed a generative diffusion emulator that "steers" climate model simulations toward historical observations to produce more accurate global temperature projections. By using a novel "condition-to-condition guidance" mechanism, the framework more than halves biases in extreme temperature predictions while preserving the essential diversity of future climate scenarios.

What Problem Does This Paper Address?

Climate risk assessment relies on understanding daily temperature distributions, especially rare and widespread extremes. However, we face a fundamental "bias-variance" trade-off in climate modeling:

  1. Climate Models (CMIP6) are our only tools for simulating future physics, but they suffer from systematic biases (e.g., overestimating historical extreme heat area by 3x).
  2. Observations (ERA5) provide the ground truth but don't exist for the future.
  3. Traditional Bias Correction often operates on individual grid cells, which can break spatial coherence and "over-correct," potentially erasing the meaningful differences between different models that represent our genuine uncertainty about the future.

The authors seek a way to blend the physical consistency of multi-model ensembles with the accuracy of historical observations without losing the "spread" of possible future outcomes.

Methods

The core of this work is a Conditional Diffusion Emulator. Unlike standard models that output a single deterministic value, this is a generative AI model (specifically a UNet-based denoiser) trained to produce entire global daily temperature fields.

The model is conditioned on four variables: model identity, day of the year, calendar year, and $CO_2$-equivalent concentration. The "secret sauce" is a technique called Condition-to-Condition Guidance:

  • Geometric Mean Sampling: At inference time, the model doesn't just sample from a specific climate model. Instead, it samples from a weighted geometric mean of a climate model's distribution and the observational distribution.
  • Steering: This acts as a "conjunction"—the model generates fields that are likely under both the physics-based simulation and the historical observations.
  • Guidance Schedule: The authors use "perfect-model" experiments (treating one model as the "truth" and others as the "source") to determine how much guidance to apply. They discovered that the influence of observations should decrease as we project further into the future, allowing the simulated dynamics to take over as forcing and lead times grow.
Guidance balances observational fidelity and model-based extrapolation.

Datasets

  • CMIP6 Simulations: Daily near-surface temperature fields from 17 different models under three Shared Socioeconomic Pathway (SSP) scenarios (SSP2-4.5, SSP3-7.0, and SSP5-8.5) from 1979 to 2100.
  • ERA5 Reanalysis: The observational reference for the period 1979–2024.
  • CanESM5 Grid: All data was remapped to a common $64 \times 128$ Gaussian grid (~2.8° resolution).

Key Insights and Ideas

  • The "Aha" Moment for Guidance: The authors found that a constant correction factor is insufficient. By using a "perfect-model" setup, they proved that the optimal guidance scale ($\gamma$) decays over time. Near the present, we should trust observations almost entirely ($\gamma \approx 0.94$); by 2100, we should rely more on the models ($\gamma \approx 0.77$).
  • Spatial Coherence Matters: Because the diffusion model generates the entire global field at once, it preserves the temperature gradients between neighboring cells. Traditional per-cell bias corrections often distort these relationships, leading to "speckled" or unrealistic local gradients.
  • Efficiency as an Enabler: The emulator can generate hundreds of thousands of daily samples far more cheaply than running a full climate model. This allows researchers to resolve "rare tail events" (like 1-in-100-year heatwaves) that are often undersampled in standard climate archives.

Key Results

  • Bias Reduction: The guided emulator more than halved median biases in the 1st and 99th percentiles of daily temperatures across 17 synthetic targets.
  • Improved Trajectories: Median annual global-mean temperature RMSE was reduced by 32–57% across the three emission scenarios.
  • Correcting Extreme Heat Area: Raw CMIP6 models historically overestimate the land area affected by extreme heat by roughly three times. The guided emulator successfully aligns these projections with ERA5 observations.
  • Future Risk: Under the high-emissions SSP3-7.0 scenario, the model projects that concurrent heat extremes will affect nearly half of all global land (46.9%) on a typical day by the 2080s.
Guided projections constrain future warming and recover a known synthetic future.

Limitations and Future Work

  • Temporal Independence: The current model draws days independently. This means it cannot directly estimate the duration of multi-day heatwaves or the risks of "temporally compound" events.
  • Resolution: The analysis is limited to moderate spatial resolution. Future work could extend this to higher-resolution, multivariate fields (e.g., including humidity for heat stress).
  • Shared Biases: The framework cannot correct for biases that are shared across all models in the CMIP6 ensemble, as it relies on the diversity of the training set to find the "correct" path.
  • Calibration: While guidance improves reliability, the authors note that uncertainty bands are still somewhat under-dispersive (covering the truth in ~70% of years vs. the nominal 90%). Future work may involve "calibration-aware" guidance.

Summary updated

Abstract

Assessing climate risks requires projections of daily temperature distributions, including rare and spatially widespread extremes. Climate simulations are costly and biased, while correcting them toward observations risks suppressing inter-model differences. We train a single machine-learning emulator, a conditional diffusion model, on seventeen CMIP6 models and ERA5 reanalysis to generate spatially coherent daily temperature fields conditioned on model identity, date, and CO2-equivalent concentration. At sampling time, diffusion guidance reduces model biases by steering each emulated climate-model distribution toward observations, with its strength selected through perfect-model experiments using withheld simulated futures. Observational influence decreases with projection horizon and forcing. Across seventeen synthetic targets, guidance more than halves median biases in the 1st and 99th percentiles of daily temperature, improves probabilistic scores and reliability, and reduces median annual global-mean temperature RMSE by 32-57% across three emission scenarios. In a detailed benchmark, guidance also outperforms established bias corrections on daily-field probabilistic scores. Efficient sampling generates large ensembles without further climate simulations. For example, our framework corrects CMIP6's roughly threefold historical overestimate of extreme-heat coverage, yet projects that concurrent extremes will affect nearly half of global land on a typical day in the 2080s under SSP3-7.0.
BibTeX
@misc{immorlano2026guiding,
  author = {Francesco Immorlano and Elijah Tavares and Felix Draxler and Juan Nathaniel and Violette Launeau and Duncan Watson-Parris and Padhraic Smyth and Pierre Gentine and Stephan Mandt},
  title = {{Guiding future global temperature distributions with simulations and observations}},
  year = {2026},
  eprint = {2611.10882},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.10882},
}
Figure