AIFS-COMPO: A Global Data-Driven Atmospheric Composition Forecasting System

Paula Harder; Johannes Flemming; Mihai Alexe; Gert Mertes; Baudouin Raoult; Matthew Chantry

physics.ao-phMar 29, 2026

AIFS-COMPO: A Global Data-Driven Atmospheric Composition Forecasting System

Authors: Paula Harder, Johannes Flemming, Mihai Alexe, Gert Mertes, Baudouin Raoult, Matthew Chantry

Organizations: European Centre for Medium-Range Weather Forecasts (ECMWF)

Abstract

We introduce AIFS-COMPO, a skilful medium-range data-driven global forecasting system for aerosols and reactive gases. Building on the ECMWF Artificial Intelligence Forecast System (AIFS), AIFS-COMPO employs a transformer-based encoder-processor-decoder architecture to jointly model meteorological and atmospheric composition variables. The model is trained on Copernicus Atmosphere Monitoring Service (CAMS) reanalysis, analysis, and forecast data to learn the coupled dynamics of weather, emissions, transport, and atmospheric chemistry. We evaluate AIFS-COMPO against a range of atmospheric composition observations and compare its performance with the operational CAMS global forecasting system IFS-COMPO. The results show that AIFS-COMPO achieves comparable or improved forecast skill for several key species while requiring only a fraction of the computational resources. Furthermore, the efficiency of the approach enables forecasts beyond the current operational horizon, demonstrating the potential of AI-based systems for fast and accurate global atmospheric composition prediction.

Figures & tables

Type	Field	Leveltype	Input/Output
NWP variables	Geopential, horizontal and vertical wind components	Pressure level	Both
	specific humidity, wind speed, temperature
NWP variables	Surface pressure, mean sea-level, 2m temperature	Surface	Both
	skin temperature, 2m dewpoint temperature
	10 m horizontal and vertical wind components
	total column water and water vapour

Table 1: Variables used in AIFS-COMPO. Pressure levels used here are 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, 1000.

Stage	Data	Years	Iterations	Learning rate range	Rollout
1. Pretraining	EAC4	2003–2023	250k	$5\cdot 10^{-4}$ – $3\cdot 10^{-7}$	1
2. Operational finetuning	Operational	2019–2023	60k	$3\cdot 10^{-4}$ – $3\cdot 10^{-7}$	1
3. Rollout finetuning	Operational	2019–2023	$1.6k\cdot 24$	$1.46\cdot 10^{-4}$ – $3\cdot 10^{-7}$	1–24

Table 2: Training stages for AIFS-COMPO.

Figure 1: A random sample of the day 3 forecast of AIFS-COMPO and IFS-COMPO for total AOD at 550nm.

Figure 2: AOD prediction performance of IFS-COMPO (red) and AIFS-COMPO (blue) compared against Aeronet observations (yellow). Top left: global RMSE, top right: global bias, bottom left: global temporal correlation, bottom right: a JJA AOD timeseries for North America.

Figure 3: Comparison of PM predictions for North America, Europe, and China. First row is showing PM 2.5 , second row is showing PM 10 .

Figure 4: Comparison of NO 2 , SO 2 , ozone, and CO predictions of AIFS-COMPO (blue) and IFS-COMPO (red) against observations in North America (first column), Europe (second column), and China (third column).

Figure 5: Evaluation of ozone profiles. Left: locations of Antarctic stations (top) and North American/European stations (bottom) in white circles. Middle and right: comparison of model predictions, AIFS-COMPO (blue) and IFS-COMP (red) with sonde observations (black). Middle: Using Antarctic stations to capture the ozone hole in September and October, plotted with logarithmic pressure scale. Right: Using all stations of North American/European over a whole year and a linear scale.

Figure 6: Ozone hole over the Southern Hemisphere (August–December 2024), showing analysis, 5-day forecasts from IFS-COMPO and AIFS-COMPO, and a 10-day forecast from AIFS-COMPO.

Figure 7: RMSE (top) and temporal correlation (bottom) for AOD (left) and PM10 (right) compared against observations. It shows three different versions of AIFS-COMPO, each after a different training stage reanalysis pretraining (ra), opertional finetuning (op) and rollout finetuning (ro).

Figure 8: RMSE of IFS-COMPO and AIFS-COMPO forecasts evaluated against CAMS analysis for total AOD at 550 nm and its components: black carbon (BC), dust, organic matter (OM), sulphate, and sea salt.

Figure 9: RMSE of IFS-COMPO and AIFS-COMPO forecasts evaluated against CAMS analysis for PM 1 , PM 2.5 , and PM 10 .

Figure 10: RMSE of IFS-COMPO and AIFS-COMPO forecasts evaluated against CAMS analysis for total column ozone (O 3 ), CO, NO 2 , and SO 2 .

Figure 11: Relative RMSE of IFS-COMPO and AIFS-COMPO forecasts evaluated against CAMS analysis for pressure-level predictions of CO, ozone, NO 2 , and SO 2 .

Air Quality Prediction

Air quality prediction estimates future pollutant levels and maps local exposure to support health warnings, planning, and pollution control.

See all 42 papers

Toward Mechanistic Interpretability of an AI Foundation Model Fine-Tuned for Atmospheric Chemistry
Jul 22, 2026Jason Y. Hu, Ivan Higuera-Mendieta, Patrick Obin Sturm +1Air Quality Prediction Mechanistic Interpretability
Stanford University, Department of Computer Science, Stanford, CA, USA. · Stanford University, Department of Earth System Science, Stanford, CA, USA · University of Southern California, Department of Earth Sciences, Los Angeles, CA, USA +1
AirMPA: A Meteorology-to-Pollution Adapter for Global Air Quality Forecasting
Oct 19, 2026Yiheng Wang, Kai Zheng, Yuetan Lin +5Air Quality Prediction Spatiotemporal Forecasting
Fudan University · Shanghai Academy of AI for Science · University of Shanghai for Science and Technology
AIFS-SUBS: Extending Data-Driven Forecasting to Sub-Seasonal Timescales
Jul 6, 2026Jakob Schloer, Steffen Tietsche, Christopher D. Roberts +6Probabilistic Forecasting Weather Forecasting
1European Centre for Medium-Range Weather Forecasts (ECMWF)

Jul 22, 2026J/K move · Enter open · S save

Toward Mechanistic Interpretability of an AI Foundation Model Fine-Tuned for Atmospheric Chemistry

Jason Y. Hu, Ivan Higuera-Mendieta, Patrick Obin Sturm, Makoto M. Kelp

Stanford University, Department of Computer Science, Stanford, CA, USA. · Stanford University, Department of Earth System Science, Stanford, CA, USA · University of Southern California, Department of Earth Sciences, Los Angeles, CA, USA+1

Air Quality Prediction Mechanistic Interpretability

Open paper AskPDF arXiv

Weather forecasting foundation models (FMs) are increasingly fine-tuned to predict air quality, offering fast global pollution forecasts at lower computational cost than conventional chemical transport models. These FMs are typically trained on reanalysis data and generate forecasts through autoregressive rollout. They do not explicitly represent governing physical or chemical processes. Therefore, high forecast skill does not reveal whether a model has learned physical mechanisms or exploits statistical regularities in its training data. Here, we present the first study of what a FM fine-tuned for atmospheric chemistry has learned by examining Microsoft's Aurora model. We impose controlled chemical perturbations on its forecasts and test them against known photochemical relationships. We then examine the internal representations that generate these forecasts. We find that Aurora captures a first-order ozone response to reactive nitrogen but does not enforce the chemical constraints that a process-based model encodes. It generates chemically inconsistent combinations of related species and relaxes localized emission features such as wildfire plumes toward background. Internally, its representations remain largely organized around the meteorology inherited during pretraining, with little structure specific to chemistry. Using sparse autoencoders, we identify internal components that causally control the chemical forecast but do not map cleanly onto individual atmospheric processes. This work provides a framework for testing whether AI forecasting systems learn atmospheric chemistry from reanalysis data. As these models are increasingly positioned to inform environmental policy decisions, we argue that composition forecasts should also be judged by their internal mechanisms rather than by benchmark skill alone.

Abstract

Air Quality Prediction

Explore similar work

Toward Mechanistic Interpretability of an AI Foundation Model Fine-Tuned for Atmospheric Chemistry

AirMPA: A Meteorology-to-Pollution Adapter for Global Air Quality Forecasting

AIFS-SUBS: Extending Data-Driven Forecasting to Sub-Seasonal Timescales