AI Just Simulated 50 Years of Climate Chemistry in a Day

50 years of atmospheric chemistry simulation — once a months-long computational marathon — can now be run in under a day, with 23 of 27 key climate variables predicted more accurately than ever before by an AI model. This isn’t a faster supercomputer. It’s a foundation model, fine-tuned to capture not just weather, but the invisible chemistry shaping our climate: ozone, carbon monoxide, aerosols, and their feedbacks with radiation and temperature. The model, called AuroraUKCA, reproduces daily atmospheric states with an average accuracy (R²) above 0.85 across the troposphere and stratosphere, matches ozone and CO levels within 4% of the original simulation, and tracks surface radiation to within 6%. Most strikingly, it runs ~100× faster than the traditional chemistry-climate model it was trained on — and remains stable over half-century-long simulations. This leap doesn’t replace physics-based models. Instead, it opens a new path: using AI not just to forecast weather, but to explore how emissions reshape air quality and climate, at a fraction of the cost and time.

The Science

Atmospheric chemistry is the silent engine of climate change. When we emit nitrogen oxides (NOx) or carbon monoxide (CO), they don’t just pollute the air — they trigger cascading reactions that alter ozone levels, affect methane lifetime, and shift the planet’s energy balance. Modeling these interactions is essential for climate policy, yet it’s also the most computationally expensive part of climate simulations. Traditional chemistry-climate models (CCMs) like the UK Chemistry and Aerosols (UKCA) model must solve thousands of stiff chemical equations at every grid point, consuming up to 90% of total runtime (Eastham et al., 2018; Long et al., 2015). This cost forces trade-offs: simplified chemistry, fewer ensemble runs, or prescribed rather than interactive components.

Enter AI foundation models. These are large neural networks pretrained on vast datasets — in this case, weather and climate data — and then fine-tuned for specific tasks. Models like Pangu-Weather and GraphCast have already matched or surpassed traditional numerical weather prediction (Bi et al., 2023; Lam et al., 2023). But until now, none could simulate interactive chemistry: the two-way feedback between emissions, atmospheric composition, and radiation.

James Weber and Makoto Kelp (2026) bridge that gap by adapting the Aurora Earth system foundation model (Bodnar et al., 2025). Aurora was originally trained on ERA5 reanalysis, CMIP6 simulations, and operational weather data for forecasting. The authors fine-tuned it on output from UKCA, a full chemistry-climate model, to predict 27 atmospheric fields — including meteorology, composition, aerosols, and radiative fluxes — across 20 pressure levels from the surface to the upper stratosphere.

The key innovation lies in how they adapted Aurora to chemistry. Trace gas concentrations vary over orders of magnitude — OH radicals can differ by 10⁴ between clean and polluted air — and must remain non-negative. Standard normalization fails here. So the authors introduced a learnable log-linear input transform (Eq. 1):

$$f(z)=w_1\min (\max (z,0),2.5)+w_2\frac{\log (\max (z,10^{-4}))-\log (10^{-4})}{-\log (10^{-4})}$$

This allows the model to blend linear and logarithmic representations per species, optimized during training. They also redesigned the loss function to balance fast- and slow-reacting species. Instead of normalizing by climatological variance (as in weather models), they normalized by the root mean square of day-to-day changes (Eq. 2–3), ensuring that both rapidly varying OH and long-lived methane contribute equally to learning. A separate conservation penalty (Eq. 4) enforces approximate mass balance for long-lived tracers, mimicking a core physical constraint absent in pure weather models.

Three model variants were developed:

• AuroraUKCA: Trained on UKCA’s control simulation to reproduce daily atmospheric states.
• AuroraUKCA,emiss: Further fine-tuned on perturbed emission scenarios to learn chemical responses.
• Auroraobs: Trained with satellite ozone data to correct model biases.

Training used single-step prediction with autoregressive evaluation — the model predicts one day ahead, then feeds that prediction back as input, repeating over decades. This mirrors how traditional models integrate forward in time, but at a fraction of the cost.

What They Found

The results are striking. In single-step prediction over 358 days in 2014, AuroraUKCA achieves an R² above 0.85 for 23 of 27 predicted fields across both troposphere and stratosphere. Ozone, CO, and temperature are reproduced with annual mean biases below 4%. Downwelling shortwave and longwave radiation at the surface are within 6% of UKCA.

Model Performance Across Atmospheric Variables

Label	Value
Ozone	0.92
CO	0.89
Temperature	0.95
OH	0.78
Downwelling SW	0.91
Downwelling LW	0.88

Single-step prediction accuracy (R²) and bias for key variables. OH shows lower R² due to extreme variability but remains physically plausible.

**

Model Performance Across Atmospheric Variables

Label	Value
Ozone	0.92
CO	0.89
Temperature	0.95
OH	0.78
Downwelling SW	0.91
Downwelling LW	0.88

Single-step prediction accuracy (R²) and bias for key variables. OH shows lower R² due to extreme variability but remains physically plausible.

Model Performance Across Atmospheric Variables**

Variable	R² (Troposphere)	R² (Stratosphere)	Annual Mean Bias
Ozone	0.92	0.94	3.1%
CO	0.89	0.87	3.8%
Temperature	0.95	0.93	2.9%
OH	0.78	0.81	12.4%
Downwelling SW	0.91	—	5.2%
Downwelling LW	0.88	—	5.7%

Data from single-step evaluation in 2014. OH shows lower R² due to extreme spatial variability but remains within physical bounds. Radiation fields are surface-only.

Perhaps most impressive is stability. When run autoregressively for 50 years — fed only initial conditions and repeating the 2014 forcing cycle — the model maintains a stable climate state without drift.

shows zonal mean Hovmöller diagrams for ozone, OH, CO, and outgoing longwave radiation. The AI model reproduces the annual cycle five times with minimal deviation, despite accumulating predictions over 18,250 days. Percent differences average under 10% for most fields, with larger errors only in polar regions during winter, where data sparsity and chemical quiescence challenge both models.

The model also learns physical responses to emissions. Trained on paired simulations with scaled NO, CO, and black carbon (BC) emissions, AuroraUKCA,emiss captures the nonlinear chemistry of tropospheric ozone. For example, when NO emissions increase, ozone rises in polluted regions (NO-limited regime) but falls in remote areas (VOC-limited regime) — a pattern the model reproduces with high fidelity (

). Similarly, it captures the radiative impact of BC: a 3× increase in emissions leads to a 15–25 W/m² reduction in surface downwelling shortwave radiation due to aerosol dimming, closely matching UKCA (

).

Even more transformative is the model’s ability to learn from observations. By adding a sparse auxiliary loss against NASA’s Aura Microwave Limb Sounder (MLS) ozone retrievals — without using observations during inference — the authors corrected a persistent high bias in UKCA’s stratospheric ozone.

shows the result: in the 250–20 hPa layer, the zonal mean error against MLS drops by 23%. The model adjusts its internal dynamics to align with real-world data, effectively performing implicit data assimilation during training.

Why This Changes Things

This work is not a replacement for process-based models. The authors are clear: UKCA remains essential for generating the training data and for probing mechanisms. But AuroraUKCA changes what’s possible in climate science.

First, it slashes the cost of scenario exploration. Running a single 50-year chemistry-climate simulation on UKCA could take weeks. AuroraUKCA does it in hours. This opens the door to high-resolution, multi-member ensembles for policy-relevant questions: What happens if we cut methane by 30%? How do wildfire emissions interact with urban pollution? How do solar geoengineering aerosols affect ozone recovery? Questions that once required supercomputing grants can now be explored on modest hardware.

Second, it integrates observations directly into model evolution. Traditional data assimilation adjusts concentrations at each time step but doesn’t change the model’s underlying dynamics. Here, the model learns to correct its own biases during training. This could revolutionize how we calibrate Earth system models, especially for variables with sparse data like reactive halogens or organic aerosols.

Third, it demonstrates that foundation models can learn physical responses — not just correlations. Many AI models fail when pushed outside their training distribution. But AuroraUKCA,emiss generalizes to emission scalings it never saw (e.g., 1.5× NO), capturing nonlinear chemical regimes. This suggests AI can begin to encode aspects of physical law, not just memorize patterns.

Compare this to past attempts. Neural networks have replaced chemical solvers within CCMs (Kelp et al., 2022), but coupling them introduces instability. Emulators have predicted surface responses (Dewey et al., 2025), but not full 3D chemistry. AuroraUKCA goes further: it’s a self-consistent, globally resolved, long-running emulator of a full chemistry-climate system.

The implications extend beyond academia. Regulatory agencies could use such models to rapidly assess the air quality and climate impacts of proposed emission rules. Cities could simulate pollution episodes under different traffic or energy policies. And global assessments like the IPCC could run far more scenarios, reducing uncertainty in climate projections.

What’s Next

The model has limits. It was trained on a single year of emissions and SSTs, so it doesn’t capture long-term climate trends. It excludes some aerosol modes and halogen species. Clouds are simplified to a binary low-cloud indicator. And while it learns responses to emissions, it doesn’t yet simulate feedbacks like changing biogenic emissions under warming.

Three frontiers stand out. First, coupling with dynamic emissions. Right now, emissions are prescribed. The next step is to let the model predict emissions — say, from wildfires or agriculture — based on climate conditions, creating a fully end-to-end system.

Second, multi-model training. AuroraUKCA learns from UKCA, but UKCA has biases. Training on an ensemble of CCMs — or combining simulations with reanalysis and satellite data — could produce a more robust, consensus-aware foundation model.

Third, uncertainty quantification. AI models are often seen as “black boxes.” Integrating probabilistic forecasting — predicting not just a single state, but a distribution of possible outcomes — would make these tools more useful for risk assessment.

The biggest question is generalization. Can a model trained on one CCM simulate another? Can it adapt to new instruments or regions? The success with MLS suggests yes — but only if we design training to prioritize physical consistency over raw accuracy.

This isn’t the end of traditional modeling. It’s the beginning of a hybrid era: where AI accelerates exploration, and physics-based models provide truth. The future of climate science may not be bigger supercomputers, but smarter models — ones that learn not just from equations, but from data, observations, and each other.