European Centre for Medium-Range Weather Forecasts
Verified evidenceContinuous decisioningAn ML forecast can run as a supported production system alongside retained physics-based forecasting.
Engineering design · Europe / global
Executive brief
A credible production transition retains parallel baselines, observable scorecards, and rollback capacity.
AI value · Operational model skill improvement
Verified4–6% overall skill improvement; up to 12% precipitation improvement in v1.1.0
Version-to-version AIFS comparison, not universal superiority over IFS; skill varies by variable and lead time.
Before
Runs physics-based numerical weather prediction from operational analyses. → Compare products, interpret uncertainty, and issue decisions through downstream services.
After
Generates deterministic global forecasts four times daily from ECMWF analyses. → Run AIFS alongside IFS, compare performance, and retain physics-based fallback and expert interpretation.
Human boundary
AIFS produces a forecast; ECMWF operations, forecasters, and downstream authorities retain release and warning decisions.
Why it matters
An ML forecast can run as a supported production system alongside retained physics-based forecasting.
Before
How the work ran before the change.
ECMWF Integrated Forecasting System
Runs physics-based numerical weather prediction from operational analyses.
ControlIFS operational verification and 24/7 support.
ECMWF forecasters and users
Compare products, interpret uncertainty, and issue decisions through downstream services.
ControlHuman/service authority; no single forecast automatically issues a warning.
What changed
An ML forecast can run as a supported production system alongside retained physics-based forecasting.
Decision rightSelection moves from a fixed rule to the model
After
How the same work runs now.
AIFS Single
Generates deterministic global forecasts four times daily from ECMWF analyses.
ControlPhysical bounding layers, open scorecards, operational monitoring, and 24/7 support.
ECMWF operations and users
Run AIFS alongside IFS, compare performance, and retain physics-based fallback and expert interpretation.
ControlIFS remains supported; model updates correct issues such as precipitation behavior.
Exception path
Operational degradation or variable-specific weakness is handled through parallel IFS comparison, monitoring, and version updates.
Work removed
Decision authority
AIFS produces a forecast; ECMWF operations, forecasters, and downstream authorities retain release and warning decisions.
Before
ECMWF Integrated Forecasting System
Runs physics-based numerical weather prediction from operational analyses.
Control: IFS operational verification and 24/7 support.
ECMWF forecasters and users
Compare products, interpret uncertainty, and issue decisions through downstream services.
Control: Human/service authority; no single forecast automatically issues a warning.
After
AIFS Single
Generates deterministic global forecasts four times daily from ECMWF analyses.
Control: Physical bounding layers, open scorecards, operational monitoring, and 24/7 support.
ECMWF operations and users
Run AIFS alongside IFS, compare performance, and retain physics-based fallback and expert interpretation.
Control: IFS remains supported; model updates correct issues such as precipitation behavior.
Work that left the path
Human role before
Operations run physics-based forecasts and forecasters interpret them.
Human role after
Operations support parallel ML and physics systems; users interpret both.
AI roleProduces an operational deterministic machine-learned global forecast.
AIFS Single 1.0.0→4–6% overall skill improvement; up to 12% precipitation improvement in v1.1.0
v1.1.0 released 2025-08-27; peer-reviewed 2026 · 24/7 operational global forecasting alongside IFS
Version-to-version AIFS comparison, not universal superiority over IFS; skill varies by variable and lead time.
Anti-pattern
Do not collapse a version-to-version gain into a claim that AI beats physics everywhere.
Questions
Portability conditions
Reputation risk
high
Evidence and authority
Current · updated
1 peer reviewed, 1 primary; publication outcomes are verified.
Bundle 1.0.0 · reviewed 2026-08-23 · stable ID 1d67e3416e59a2dd
Google DeepMind and EMBL-EBI · Continuous decisioning
Researchers can use precomputed structure predictions, then decide what experimental validation is required.
Google Research Flood Hub · Continuous decisioning
A global model can transfer hydrology into ungauged basins.
Mayo Clinic primary-care practices · Continuous decisioning
A routine ECG can trigger evaluation for asymptomatic low ejection fraction before specialist suspicion.
Google · Continuous decisioning
Novel breakages need not wait for a comprehensive post-submit test cycle.