[KongchangAI]
· 2 min read· 1,065 words

Diffusion Models Guide Climate Simulation for Accurate Extreme Heat Distribution Prediction

Diffusion Models Guide Climate Simulation for Accurate Extreme Heat Distribution Prediction

Conditional diffusion models with diffusion guidance efficiently generate low-bias daily temperature fields and reveal future extreme heat risks.

The study proposes a machine learning climate emulator based on a conditional diffusion model, trained on 17 CMIP6 models and ERA5 reanalysis data, conditioned on model identity, date, and CO2-equivalent concentration to generate spatially coherent daily temperature fields. Diffusion guidance during sampling steers simulations toward observations, with strength set by perfect-model experiments and observational influence weakening with forecast horizon and forcing. On 17 synthetic targets, guidance more than halves median biases at the 1st and 99th percentiles of daily temperature, improves probabilistic scores and reliability, reduces median global mean temperature RMSE by 32-57% across three emission scenarios, and outperforms existing bias correction methods on daily field probabilistic scores. The framework corrects CMIP6's roughly threefold historical overestimation of extreme heat coverage and projects that under SSP3-7.0, nearly half of global land will face concurrent extreme events on a typical day in the 2080s.

Climate Simulation Challenges and Machine Learning Breakthroughs

Climate risk assessment depends on accurate predictions of daily temperature distributions, especially for rare but wide-impact extreme events. However, traditional climate simulations are extremely computationally expensive and suffer from systematic biases. Directly correcting simulation results with observational data can reduce bias but may suppress inter-model differences, thereby affecting uncertainty assessment. This study proposes a machine learning emulator based on a conditional diffusion model, designed to generate spatially coherent daily temperature fields while preserving inter-model diversity.

The model is trained on seventeen CMIP6 models and ERA5 reanalysis data, conditioned on model identity, date, and CO2-equivalent concentration. Through diffusion guidance, the simulated distribution can be steered toward observations during sampling, thereby reducing bias. The guidance strength is selected through perfect-model experiments using held-out simulated future data. The study finds that observational influence weakens with increasing forecast horizon and forcing, which aligns with physical intuition.

CMIP6 (Coupled Model Intercomparison Project Phase 6) is the core framework coordinating multi-model simulations for the global climate research community, bringing together dozens of Earth system models from dozens of institutions. Each model differs in physical parameterizations, resolution, cloud microphysics, and aerosol treatment, so outputs under the same emission scenario exhibit structural differences across models. This inter-model spread is one of the main sources of uncertainty in climate projections and is also the basis for ensemble methods (multi-model averaging or weighting). CMIP6 data is the core foundation of the IPCC Sixth Assessment Report, but its raw output often has significant biases at regional and daily scales, requiring downscaling or bias correction before it can be used for impact assessment.

ERA5 is the fifth-generation global atmospheric reanalysis dataset released by the European Centre for Medium-Range Weather Forecasts (ECMWF), generated by assimilating a numerical weather prediction model with massive historical observations (satellites, radiosondes, surface stations, etc.) to produce global three-dimensional atmospheric fields at hourly resolution and approximately 31 km resolution from 1940 to the present. Reanalysis data is not pure observation but a fusion product of model and observations; however, for daily temperature and circulation variables it is generally regarded as the most reliable "observational proxy" and is widely used as training targets for climate model evaluation and bias correction.

Conditional diffusion models are a class of generative models whose core idea is to recover the data distribution from Gaussian noise through an iterative denoising process, with additional information (such as model identity, date, CO2 concentration) available at each step to condition the generation. Compared with GANs, diffusion models have more stable training and more complete mode coverage, making them particularly suitable for generating high-dimensional spatial fields. Diffusion guidance involves injecting additional gradients or biases into the denoising network during sampling so that generated samples move toward a target distribution (here, observations) without retraining the model. Guidance strength is a key hyperparameter: too strong and it flattens inter-model differences, degenerating into observation copying; too weak and bias correction is insufficient. Therefore, the study uses perfect-model experiments to calibrate it.

Bias Correction and Probabilistic Score Improvements

Tests on seventeen synthetic targets show that guidance more than halves the median bias of the 1st and 99th percentiles of daily temperature while improving probabilistic scores and reliability. Specifically, under three emission scenarios, the median global mean temperature RMSE is reduced by 32-57%. In detailed benchmark tests, guidance also outperforms existing bias correction methods on daily field probabilistic scores. This indicates that the method can not only reduce bias but also improve prediction reliability.

Efficient sampling enables the generation of large ensembles without further climate simulations, which is crucial for risk assessment. For example, the framework corrects CMIP6's historical overestimation of extreme heat coverage by about a factor of three, but projects that under SSP3-7.0, by the 2080s, concurrent extreme events on a typical day will affect nearly half of global land. This projection highlights the severity of future extreme heat.

Probabilistic scores (such as the Continuous Ranked Probability Score, CRPS, and logarithmic score) measure how well a forecast distribution matches observed actual values, penalizing not only mean bias but also insufficient or overconfident ensemble spread. Reliability refers to whether forecast probabilities are consistent with observed frequencies: if a model assigns a 70% probability to an extreme event, such events should occur in approximately 70% of samples. Unlike traditional univariate bias correction methods such as quantile mapping, these scores, when evaluated on spatial fields, can simultaneously test spatial correlation structure and event synchrony, and therefore better reflect a method's ability to correct the coverage of extreme heatwaves.

Projections and Implications of Future Extreme Heat

This study not only provides an effective bias correction method but also reveals the potential impacts of future extreme heat. Through diffusion guidance, the model can more accurately predict the spatiotemporal distribution of extreme events, providing a scientific basis for climate adaptation and risk assessment. Although the method performs excellently on synthetic targets, its effectiveness in practical applications still requires further validation. In addition, the phenomenon that observational influence weakens with forecast horizon also suggests that model uncertainty may dominate in long-term projections.

Overall, this study demonstrates the potential of machine learning in climate simulation, particularly in reducing bias and generating large ensembles. In the future, as more observational data accumulates and models improve, this approach is expected to become an important tool for climate risk assessment.

Study Overview

One-line Summary

Conditional diffusion models combined with diffusion guidance efficiently generate low-bias daily temperature fields and reveal future extreme heat risks.

Paragraph Summary

The study proposes a machine learning climate emulator based on a conditional diffusion model, trained on 17 CMIP6 models and ERA5 reanalysis data, conditioned on model identity, date, and CO2-equivalent concentration to generate spatially coherent daily temperature fields. During sampling, diffusion guidance steers the simulated distribution toward observations, with guidance strength determined by perfect-model experiments, and observational influence weakening with increasing forecast horizon and forcing. On 17 synthetic targets, guidance more than halves the median bias of the 1st and 99th percentiles of daily temperature, improves probabilistic scores and reliability, reduces median global mean temperature RMSE by 32-57% under three emission scenarios, and outperforms existing bias correction methods on daily field probabilistic scores. The framework corrects CMIP6's historical overestimation of extreme heat coverage by about a factor of three, while projecting that under SSP3-7.0, nearly half of global land will simultaneously experience extreme events on a typical day in the 2080s.

#conditional diffusion model#climate simulation#CMIP6#diffusion guidance#extreme heat prediction#bias correction
Share:

Related articles