ECMWF AIFS-COMPO Extends Pollution Forecasts, but Still Depends on Physics

ECMWF AIFS-COMPO has extended experimental global atmospheric composition forecasts to 10 days, twice the range of the operational system it seeks to complement. Yet the AI model still needs that physics-based system to establish its starting conditions.

The European Centre for Medium-Range Weather Forecasts introduced the system on October 16, 2026. It adapts ECMWF’s Artificial Intelligence Forecast System to predict aerosols, reactive gases, and weather variables together.

That shift matters because atmospheric composition forecasting supports air-quality warnings, health planning, wildfire-smoke tracking, and analysis of events such as the Antarctic ozone hole. The operational Copernicus Atmosphere Monitoring Service currently produces five-day global forecasts twice daily.

AIFS-COMPO offers a different computational route. Its developers report competitive results for several pollutants while using far fewer computing resources. However, it remains experimental, covers fewer variables, and cannot yet assimilate observations independently.

The central contest is therefore not AI against atmospheric science. It is a faster learned forecast model against the physical machinery that created its training data and continues to prepare every forecast.

ECMWF AIFS-COMPO Turns Five-Day Forecasts Into Ten-Day Experiments

The immediate change is a longer, faster forecast, not the replacement of the operational CAMS system.

ECMWF has released public forecast charts from AIFS-COMPO, an experimental system for predicting global atmospheric composition. The charts extend to 10 days and cover major pollutants, gases, and aerosol categories.

The model produces forecasts for particulate matter, including PM2.5 and PM10. It also predicts carbon monoxide, nitrogen dioxide, ozone, sulphur dioxide, and several aerosol types.

Those aerosols include dust, sea salt, sulphate, and material associated with biomass burning. Outputs are available at the surface, on pressure levels, or as total atmospheric columns, depending on the variable.

The release followed the tenth CAMS General Assembly, held in Athens from October 13 through October 15. ECMWF announced the chart availability during the gathering, according to the CAMS model release.

The operational comparison is IFS-COMPO, a configuration of ECMWF’s Integrated Forecasting System. It represents atmospheric processes through physical and chemical equations.

IFS-COMPO models how substances move, react, enter the atmosphere, and leave it. It also assimilates satellite observations, meaning it combines measurements with model estimates to produce better starting conditions.

That system issues five-day forecasts twice each day. AIFS-COMPO extends the range to 10 days because its forecasts require less computation after training.

Longer range does not automatically mean useful accuracy at every location or forecast hour. Forecast errors generally accumulate as a model advances further from its initial state.

ECMWF says its early evaluation found comparable or improved skill for several important atmospheric variables. The supporting paper also reports better results for some aerosol and particulate-matter measures.

One comparison examined aerosol optical depth, which measures how particles reduce light passing through the atmosphere. AIFS-COMPO recorded lower error across the five days covered by both models during the reported test period.

Another case examined the 2024 Antarctic ozone hole. The AI system continued to represent the event in its day-ten forecast, beyond the operational product’s current five-day range.

These results make the experiment more than a visualization demo. They show that a learned model can retain useful large-scale atmospheric signals beyond the existing operational horizon.

However, ECMWF still describes IFS-COMPO as its benchmark for accuracy and reliability. AIFS-COMPO has not displaced that benchmark or entered the same operational role.

The distinction matters for health agencies, environmental analysts, and downstream data services. Operational products need dependable schedules, documented failure handling, and consistent performance across many conditions.

A new model can outperform selected measures without satisfying all those requirements. The 10-day release starts a broader validation process rather than ending one.

Why AI Atmospheric Composition Forecasting Is Harder Than Weather Prediction

Pollution forecasting requires a model to follow weather while also accounting for emissions, chemistry, transport, and removal processes.

Data-driven weather forecasting has already produced systems that compete with conventional numerical models. Atmospheric composition presents additional difficulties that weather variables alone do not capture.

Pollutants originate from uneven and changing sources. Wildfires, industrial activity, traffic, dust storms, vegetation, and sea spray can each alter the atmospheric state.

Their effects unfold at different spatial and temporal scales. A gas can react chemically, while particles can combine, settle, or wash out during precipitation.

Wind transports those substances across national borders. Temperature, sunlight, humidity, clouds, and precipitation can also change their concentrations or chemical behavior.

Measurements remain uneven across the planet. Weather observations are imperfect, but atmospheric composition data can be especially sparse for certain compounds and regions.

The existing CAMS production system handles these interactions through detailed physical and chemical calculations. It represents transport, emissions, chemical conversion, aerosol physics, deposition, and removal by precipitation.

It also uses satellite retrievals for aerosol optical depth, ozone, carbon monoxide, nitrogen dioxide, and sulphur dioxide. Data assimilation incorporates those observations into the estimated current atmospheric state.

AIFS-COMPO learns statistical relationships within outputs from that established system. It was pretrained using the CAMS EAC4 reanalysis covering 2003 through 2023.

A reanalysis combines historical observations with a consistent model to reconstruct past atmospheric conditions. It supplies a long, structured record when direct observations alone would be incomplete.

Developers then fine-tuned AIFS-COMPO on operational CAMS analyses and forecasts from 2019 through 2023. This stage exposed the model to a more recent production configuration.

The final training stage used progressively longer rollouts. An autoregressive rollout repeatedly feeds a model’s previous prediction into its next forecast step.

This process helps address error accumulation. A model trained only to predict one short interval can become unstable when asked to repeat that operation many times.

According to the AIFS-COMPO paper, the system advances in three-hour increments. It predicts 187 meteorological and atmospheric composition variables.

The model operates on an approximately 80-kilometer global grid. Its outputs capture broad transport patterns, but that resolution cannot directly represent every neighborhood-scale pollution source.

Its architecture uses an encoder, processor, and decoder. Graph neural-network components transfer fields between the input grid and a lower-resolution internal representation.

The processor uses a transformer-based design to learn relationships across the global state. The decoder then projects the learned representation back onto the forecast grid.

Meteorology and atmospheric composition are predicted jointly. This design lets the system model connections between wind, temperature, moisture, gases, and particles within one forecast process.

That joint approach is important. Pollution does not evolve independently from weather, and weather can also interact with aerosols and atmospheric chemistry.

The model’s breadth separates it from narrower systems that estimate one pollutant or improve local resolution. ECMWF’s researchers describe it as the first AI model with global, three-hourly forecasts across a broad composition-variable set.

That description belongs to the research team and should be read within its stated scope. Other machine-learning systems already forecast aerosols, regional chemistry, and selected global environmental variables.

Microsoft’s Aurora, for example, has demonstrated a foundation-model approach across several Earth-system forecasting tasks. Regional systems have also applied deep learning to pollution transport and atmospheric chemistry.

AIFS-COMPO’s differentiator is its integration with ECMWF’s established forecasting environment. Its training data, initial conditions, comparison model, and public charts all connect to CAMS operations.

The Speed Advantage Changes the Forecasting Equation

AIFS-COMPO shows that atmospheric composition forecasts can become dramatically cheaper after training, creating room for longer ranges and larger ensembles.

The reported computational difference is substantial. The research team says AIFS-COMPO produces a five-day global forecast in about 50 seconds on one graphics processor.

The comparison run for IFS-COMPO takes roughly 1,000 seconds across 8,000 central processing units, according to the paper. These figures describe different computing architectures, so they are not a direct hardware-cost calculation.

Even with that qualification, the operational opportunity is clear. Fast inference can support more forecast runs without repeating the full numerical calculation each time.

That capacity could enable larger ensembles. An ensemble runs multiple forecasts with varied starting conditions or model assumptions to estimate uncertainty.

Atmospheric composition services could also test alternative configurations more quickly. Researchers might examine more cases, longer horizons, or targeted extreme events within the same computing budget.

The efficiency also explains the move from five to 10 days. Conventional chemistry models become costly because each forecast must calculate many interacting processes across the globe.

A learned model performs most of its expensive optimization during training. Once trained, it applies its learned relationships to new initial states much faster.

The benefit is particularly relevant for smoke and dust that travel across continents. A longer forecast can give agencies more time to examine the possible path of a large plume.

It could also help researchers follow slow-developing stratospheric events. The reported Antarctic ozone-hole case illustrates that longer-horizon use.

However, the AI model does not receive a raw collection of current observations and create a forecast independently. Its first state comes from the operational CAMS analysis.

That analysis already incorporates satellite data using the physics-based system’s assimilation process. AIFS-COMPO therefore inherits a carefully prepared estimate of the present atmosphere.

This dependence changes the meaning of the speed comparison. The AI forecast is inexpensive after initialization, but its input required a large operational system to produce.

The relationship resembles a fast runner receiving the baton from a slower, more elaborate process. Removing the first stage would leave the runner without a reliable starting position.

AIFS-COMPO is consequently a forecast model, not yet a complete forecasting system. ECMWF explicitly notes that it does not perform data assimilation.

The operational model also produces more variables and works at higher spatial resolution. The comparison data for IFS-COMPO use an approximately 40-kilometer grid, against about 80 kilometers for AIFS-COMPO.

Resolution affects local interpretation. A broad global field can represent a transported smoke plume while missing sharper concentration differences near cities, roads, or fires.

The AI system also lacks deposition fluxes, which describe material moving from the atmosphere onto the surface. It represents fewer chemical species than IFS-COMPO.

Those omissions limit direct substitution. Users who depend on a missing variable cannot recover it from faster forecasts of a different quantity.

The meaningful near-term model is therefore hybrid. Physics-based assimilation establishes the atmospheric state, while AI performs rapid forward prediction.

This division could preserve the observation-handling strengths of established systems while reducing the cost of repeated forecasts. It also keeps a physical benchmark available for comparison.

The approach pressures conventional forecasting programs to justify where expensive simulation adds essential value. It does not make those simulations obsolete.

Instead, it moves the performance question from a single accuracy score to a broader tradeoff. Agencies must weigh speed, forecast range, resolution, variable coverage, reliability, and unusual-event behavior together.

The CAMS AI Model Still Carries Its Teacher’s Limits

AIFS-COMPO learns from a physics-based system, so matching that system can reproduce its strengths and some of its errors.

Training on CAMS data gives AIFS-COMPO two decades of consistent global fields. It also makes the model dependent on decisions embedded in that dataset.

Reanalysis is not direct observation. It is a model-guided reconstruction built by combining available measurements with physical assumptions.

Sparse observations can give the model more influence over the reconstructed state. This issue becomes important for pollutants with limited coverage or poorly characterized emissions.

The AI system can learn recurring relationships in that reconstructed history. It cannot automatically determine whether every relationship reflects nature, a modeling approximation, or a persistent data bias.

Independent observations provide the strongest test against that circularity. The research team evaluated 2024 forecasts using multiple monitoring networks for aerosols, particulate matter, reactive gases, and ozone.

That evaluation is more informative than simply measuring agreement with CAMS analysis. It asks whether predictions match instruments that were not serving as the forecast target.

Initial results were mixed in the useful scientific sense. AIFS-COMPO matched or improved upon IFS-COMPO for some variables, while remaining competitive for others.

Aerosol optical depth and particulate matter were among the stronger reported areas. Performance across every location, season, chemical species, and extreme event remains less certain.

The limited evaluation period deserves attention. One year can contain many conditions, but operational atmospheric services must perform across changing climates and rare events.

Time-varying emissions present another gap. The current AIFS-COMPO configuration does not explicitly accept them as forcing inputs.

A forcing input supplies information that directly drives future changes, such as updated wildfire emissions. Without it, a model must infer future behavior from its initial state and learned patterns.

That design can work when an event evolves like familiar examples. It becomes less dependable when emissions change suddenly after initialization.

A wildfire can intensify, a volcano can erupt, or an industrial accident can release material unexpectedly. Historical patterns alone cannot fully specify those new emissions.

ECMWF is working to integrate fire data from the Global Fire Assimilation System. The institution is also examining other emissions datasets and direct atmospheric observations.

These additions would make the model more responsive to unfolding events. They would also complicate its architecture and validation requirements.

The current system’s lower spatial resolution creates another constraint. An 80-kilometer grid can support global forecasting without resolving every urban pollution gradient.

Regional forecasting systems often need finer grids and local emissions inventories. They may also incorporate monitoring stations whose coverage varies greatly between cities.

A global AI forecast could provide boundary conditions for those systems. It should not be confused with a street-level exposure estimate.

Forecast users also need uncertainty information. A single deterministic field can look precise even when its confidence varies across regions and lead times.

The public forecast charts make experimental outputs accessible, but charts alone do not establish operational trust.

Agencies will need sustained verification against independent instruments. They will also need documented behavior during dust storms, wildfires, ozone episodes, and unexpected emissions.

The most important skeptical point is therefore not that AI ignores physics. The model absorbs physics indirectly through CAMS-generated training data and initialized analyses.

The concern is whether that indirect representation remains dependable outside familiar conditions. Rare events and rapid source changes provide the clearest tests.

ECMWF’s own positioning reflects that uncertainty. Laurence Rouil, CAMS director, said the operational system remains the benchmark for accuracy and reliability.

That acknowledgment prevents the release from becoming a premature replacement claim. It places AIFS-COMPO inside an evaluation program with specific known gaps.

Physics and AI Are Becoming One Forecasting Stack

The likely outcome is not a winner-take-all contest, but a layered system where physical models, observations, and AI handle different tasks.

Physics-based models retain several structural advantages. Their equations explicitly represent mechanisms, accept changing emissions, and can support variables absent from learned systems.

They also provide the analyses used to initialize AIFS-COMPO. That role will remain central until an AI system can assimilate heterogeneous observations with comparable reliability.

AI models offer a different advantage. They can execute forecasts quickly enough to expand range, increase ensemble size, and shorten experimentation cycles.

A layered forecasting stack could combine these properties. A physical system would assimilate observations and maintain a comprehensive atmospheric state.

One or more AI models could then produce rapid forecasts from that state. Numerical forecasts would continue as references, fallbacks, and sources for variables outside the AI model.

Specialized regional models could add local detail. Human forecasters and environmental agencies would interpret the combined evidence for public decisions.

This arrangement also creates productive competition. If an AI model repeatedly matches a costly calculation, researchers can examine which parts of that calculation remain necessary.

If it fails during an unusual event, the failure can expose missing inputs or weakly represented processes. Both outcomes contribute to model development.

AIFS-COMPO enters a growing field of learned Earth-system models. Weather-focused systems have already shifted expectations about computational efficiency.

Atmospheric composition expands that test into a domain with complex chemistry and less complete observation coverage. Success here would support AI forecasting beyond standard weather variables.

The consequences could extend to climate monitoring, greenhouse-gas analysis, and emergency response. Each field would still require its own validation and input design.

The system could also broaden access to atmospheric forecasting research. A model that runs on one GPU is easier to test than a global chemistry simulation requiring thousands of processors.

Access remains incomplete, however. ECMWF says underlying AIFS-COMPO data will become available after further validation and maturation.

Public charts let users inspect outputs, but researchers need downloadable fields for independent evaluation. They also need documentation covering model versions, training data, and known limitations.

The use of Anemoi forecasting tools offers one path toward reproducibility. Anemoi is an open-source framework co-developed by ECMWF and national meteorological services.

Open tooling does not automatically make a trained model or its complete dataset available. Still, it can help other institutions study the broader architecture.

The competitive pressure therefore falls on both sides. Conventional models must demonstrate where their added computation improves decisions.

AI developers must show reliability beyond selected benchmarks. They must also explain dependencies that are easy to hide behind fast inference figures.

For public agencies, a cheaper forecast is valuable only if its quality is measurable. Longer range matters only when users understand how confidence changes over time.

AIFS-COMPO makes those questions operational rather than theoretical. Its charts now give forecasters and researchers a live experimental product to follow.

What to Watch as AIFS-COMPO Moves Beyond the Demo

Three signals will determine whether AIFS-COMPO becomes an operational component: independent validation, live emissions, and usable data access.

The first signal is sustained verification across specific atmospheric events. ECMWF has already cited aerosol, particulate-matter, and ozone results from its initial evaluation.

The next evidence should cover more seasons and unusual episodes. Large wildfires, volcanic plumes, dust outbreaks, urban pollution events, and rapid stratospheric changes deserve separate analysis.

A model that performs well on routine conditions can still fail when new emissions dominate the forecast. Event-level verification would strengthen the case for operational use.

Poor or inconsistent results would weaken replacement claims without eliminating the model’s research value. They could reveal where hybrid correction or specialized inputs are needed.

The second signal is the integration of time-varying fire and other emissions. ECMWF says work is underway to add information from its fire assimilation system.

This change would address one of the current model’s clearest limitations. It would let future emissions influence forecasts rather than leaving the model to infer them from prior states.

The important test will be whether new inputs improve rapidly evolving events without degrading routine performance. Researchers should also explain how missing or delayed emissions data affect forecasts.

The third signal is the release of underlying forecast data and fuller evaluation tools. Downloadable outputs would allow universities, agencies, and independent scientists to test regional performance.

External validation can identify weaknesses that global averages conceal. It can also show whether 10-day forecasts remain useful for specific pollutants and locations.

Operational status would require more than public data. Users should watch for documented service schedules, stable model versions, uncertainty products, and integration into established CAMS workflows.

The forecast horizon itself is not the final score. A reliable day-six warning can matter greatly, while an unstable day-ten map can create false confidence.

AIFS-COMPO’s real contribution is a credible demonstration that global atmospheric composition prediction can use a much faster computational path. Its limits reveal why the existing system still matters.

Researchers, public-health agencies, and environmental data teams should compare the experimental charts with local observations and operational CAMS forecasts. Focus on difficult events rather than visually convincing routine maps.

That comparison will answer the decisive question: can ECMWF AIFS-COMPO preserve useful skill when emissions change quickly and the atmosphere behaves unlike its training history?