Source-linked AI summary
AIFS -- ECMWF's data-driven forecasting system
Simon Lang, Mihai Alexe, Matthew Chantry, Jesper Dramsch, Florian Pinault, Baudouin Raoult, Mariana C. A. Clare, Christian Lessig, Michael Maier-Gerber, Linus Magnusson, Zied Ben Bouallègue, Ana Prieto Nemesio, Peter D. Dueben, Andrew Brown, Florian Pappenberger, Florence Rabier
TL;DR
Accurate medium-range global weather forecasting has motivated data-driven alternatives to physics-based NWP, but operational systems require suitable models and pipelines. This paper introduces ECMWF’s AIFS, a graph-and-transformer forecasting system trained on reanalysis and operational analyses. AIFS delivers highly competitive upper-air and surface forecasts, accurate tropical cyclone tracks, and public forecasts run four times daily.
Problem
Accurate medium-range global weather forecasting is important, while data-driven models require operational forecasting systems built around historical and operational weather analyses.
Method
AIFS combines GNN encoder and decoder blocks with a transformer processor, weighted-MSE training, and parallelism for high-resolution forecasting.
Results
AIFS produces highly competitive forecasts for upper-air and surface variables, remains strong against analyses and observations, and produces accurate tropical cyclone track forecasts.
Takeaways & Limitations
AIFS forecasts run four times daily at ECMWF and are available publicly under ECMWF’s open data policy.
Takeaways & Limitations
Weighted-MSE training can blur forecast fields at longer lead times, while height-based loss scaling reduces stratospheric forecast skill.
Abstract
from arXiv · showhide
Machine learning-based weather forecasting models have quickly emerged as a promising methodology for accurate medium-range global weather forecasting. Here, we introduce the Artificial Intelligence Forecasting System (AIFS), a data driven forecast model developed by the European Centre for Medium-Range Weather Forecasts (ECMWF). AIFS is based on a graph neural network (GNN) encoder and decoder, and a sliding window transformer processor, and is trained on ECMWF's ERA5 re-analysis and ECMWF's operational numerical weather prediction (NWP) analyses. It has a flexible and modular design and supports several levels of parallelism to enable training on high-resolution input data. AIFS forecast skill is assessed by comparing its forecasts to NWP analyses and direct observational data. We show that AIFS produces highly skilled forecasts for upper-air variables, surface weather parameters and tropical cyclone tracks. AIFS is run four times daily alongside ECMWF's physics-based NWP model and forecasts are available to the public under ECMWF's open data policy.
1 Introduction
ECMWF developed AIFS as an operationally ready data-driven forecasting system, building on rapid progress in machine-learning weather prediction and ERA5-based training.
- Data-driven weather models have outperformed leading physics-based NWP models on standard scores including 500 hPa geopotential RMSE and ACC.
- These models are generally trained on historical ERA5 reanalysis data and initialized for real-time forecasts with operational NWP analyses.
- ECMWF implemented an experimental data-driven forecasting system designed for operational use.
- The system includes an end-to-end pipeline for dataset generation, reproducible training, operational inference, verification, product generation, and forecast dissemination.
- ECMWF’s model is called AIFS, the Artificial Intelligence Forecasting System, and is developed alongside the Anemoi toolbox for data-driven weather forecasting.
2 Model
AIFS combines graph-based encoding and decoding with a transformer processor operating on a reduced Gaussian grid. Its modular parallel design supports high-resolution forecasting and future probabilistic extensions.
- Architecture: AIFS uses an encoder-processor-decoder architecture in which graph networks map ERA5 data to a latent processor grid and back.The encoder aggregates information onto the processor grid, while the decoder projects latent states back to the input grid.
- Architecture: The updated model uses attention-based GNNs in the encoder and decoder and a pre-norm transformer with shifted-window attention in the processor.
- Grid design: The encoder and decoder operate at native ERA5 resolution on the approximately 31 km N320 reduced Gaussian grid.Reduced Gaussian grids provide more uniform global resolution and fewer grid points and edges than comparable regular longitude-latitude grids.
- Processor: Attention windows restrict local processor interactions while allowing information to travel across the grid through multiple layers.The illustration shows propagation within 6 layers, whereas AIFS uses 16 processor layers.
- Features: AIFS adds eight learnable features to nodes and graph edges so the network can learn information absent from inputs or predefined features.
- Parallelism: Sequence and tensor parallelism split one model instance across GPUs, enabling training on high-resolution data and multiple autoregressive steps.Attention heads can be sharded across GPUs, while encoder and decoder nodes and edges can alternatively be sharded by one-hop graph neighborhoods.
- Parallelism: AIFS scales quasi-linearly to at least 2048 GPUs, distributing computation and activations to reduce runtime and increase available model memory.
- Probabilistic extension: Device-to-device communication also supports ensemble-based training toward probabilistic forecast scores and larger ensembles.
3 Training
AIFS learns six-hour atmospheric-state forecasts from preceding analyses and extends them through autoregressive rollouts. Training combines ERA5 pre-training, rollout optimization, weighted MSE, and distributed GPU execution.
- Forecast task: AIFS receives atmospheric states at t−6h and t0 and predicts the state at t+6h.Table 1 lists the model’s input and output variables.
- Forecast task: Longer forecasts are generated autoregressively by feeding the model its own predictions, with rollout training extended to 72 h.
- Training schedule: Pre-training uses ERA5 from 1979–2020 for 260,000 steps, followed by ERA5 rollout training on 1979–2018 with up to 12 autoregressive steps.
- Objective: The objective is an area-weighted MSE with empirical per-variable scaling, reduced vertical-velocity weight, and linearly decreasing weights with height.Upper-atmosphere levels such as 50 hPa contribute relatively little to the total loss.
- Compute: Training uses distributed data parallelism, mixed precision, and 64 GPUs, taking about one week for the full process.A 10-day forecast takes approximately 2 minutes 30 seconds on one A100 GPU, including forecast-data input and output.
4 Results
AIFS shows strong forecast skill across upper-air variables, surface weather, precipitation, and tropical cyclone tracks, with more than 12 hours’ advantage over IFS at longer lead times. Verification against analyses and observations reveals both broad improvements and limitations, including precipitation variability and increasing forecast-field smoothing.
- Upper-air forecasts: More than 12 h of forecast advantage is reported for AIFS over IFS at longer lead-times in Northern Hemisphere 500 hPa geopotential ACC.Figure 4 compares AIFS, IFS, and ERA5 using 2-, 6-, and 10-day forecasts with 30-day running means.
- Verification: AIFS is evaluated against operational analyses and direct observations, including radiosondes and SYNOP measurements across upper-air and surface variables.The verification includes ACC, RMSE, forecast activity, and observational comparisons for geopotential, temperature, wind, and precipitation.
- Upper-air forecasts: AIFS improves upper-air scores throughout the troposphere up to 100 hPa, with improvements of approximately 10% over IFS.IFS performs better at 50 hPa, and AIFS is generally better after day 1, while day-1 analysis verification can favor IFS.
- Verification: At 100 hPa, AIFS temperature RMSE and ACC appear worse than IFS against analyses but improved against Northern Hemisphere extra-tropical radiosonde observations.This contrast shows that the verification reference affects the apparent ranking of the systems.
- Surface forecasts: AIFS surface forecasts consistently outperform IFS for variables including 2 m temperature and 10 m wind when verified against analyses and observations.The comparison also indicates that higher spatial resolution is especially important for surface-field performance.
- Precipitation: Total-precipitation skill is mixed: AIFS is worse than IFS at shorter extra-tropical ranges, better at longer ranges, and generally improved in the tropics.The cited metric for the extra-tropics is SEEPS.
- Forecast-field structure: AIFS forecast activity is somewhat reduced relative to IFS, with small-scale structures increasingly washed out as lead time grows.The reduction is less pronounced than for the IFS ensemble-mean forecasts, which filter unpredictable features with lead time.
- Tropical cyclones: AIFS has lower tropical-cyclone position errors than IFS, but larger intensity errors and an indication of fewer generated cyclones.The lower position errors are associated with reduced slow bias in cyclone propagation, while lower mean intensity is linked to ERA5 and IFS analysis intensity and forecast smoothing.
5 Discussion and conclusions
AIFS delivers highly competitive upper-air, surface, and tropical-cyclone forecasts, while remaining publicly available and operationally scalable. Its main limitations are smoothing at longer lead times, reduced stratospheric skill, reduced intensity in some high-impact systems, and unresolved fine-tuning and loss-design choices.
- AIFS produces highly competitive forecasts for upper-air and surface variables, verified against both NWP analyses and radiosonde and SYNOP observations, and accurate tropical cyclone tracks.
- Weighted-MSE training causes forecast fields to become increasingly smoothed or blurred at longer lead times, reducing small-scale structure and forecast activity.Preliminary results indicate that probabilistic objectives can produce sharper forecast fields throughout the forecast.
- AIFS has reduced stratospheric forecast skill because loss weights decrease linearly with height, and some high-impact systems such as tropical cyclones have reduced intensity.The paper identifies improved height-dependent loss scaling, higher-resolution reanalyses, and refined or probabilistic objectives as potential improvement paths.
- Further work is needed to assess data-driven models in practice and compare them with physics-based NWP models, while AIFS fine-tuning on operational IFS analyses remains relatively ad hoc.
- AIFS’s modular architecture supports flexibility, scalability, extensibility, and prospective large-member ensemble forecasts because individual data-driven forecasts are comparatively inexpensive.
- AIFS is operationally run four times daily at ECMWF, with forecasts available publicly under ECMWF’s open data policy.