ECMWF takes AIFS-ENS operational as ensemble AI forecasts

On July 1, 2025, ECMWF moved the ensemble version of its Artificial Intelligence Forecasting System into operations. The system launched as AIFS ENS v1, and it runs side by side with the traditional physics-based IFS ensemble.

Current GribStream note: this post records the July 2025 operational launch. ECMWF later moved AIFS ENS to v2 at the 2026-05-12 06 UTC run. GribStream keeps that history in the same AIFS Ensemble dataset; see AIFS v2 fields are now visible on GribStream for the May 2026 update.

What is AIFS-ENS

AIFS-ENS is ECMWF's AI-based probabilistic global forecast. It uses a stochastic ML model trained on reanalysis and fine-tuned on operational analyses, then initializes member-for-member from the IFS ensemble.

ECMWF's operational v1 configuration provided:

• 51 members (50 perturbed + 1 control).
• 00/06/12/18 UTC cycles.
• 6-hourly steps out to 360 h (15 days).
• 0.25 deg global grid (open-data subset).
• N320 / about 31 km model resolution before open-data interpolation.
• 13 pressure levels from 50 hPa to 1000 hPa for atmospheric pressure-level fields.

One detail is worth calling out: ECMWF says the AIFS ENS control member is not the same kind of unperturbed baseline as the control in a physics-based ensemble. It starts from the unperturbed IFS ENS control initial condition, but the AIFS ENS model itself remains stochastic.

Why it matters

Operational AI ensembles make it possible to compare AI vs. physics guidance using similar horizons and a consistent ensemble structure. That is valuable for:

• Uncertainty-aware planning using probabilistic thresholds.
• Benchmarking AI guidance against the IFS ensemble.
• Scenario analysis for energy, aviation, marine, and hydrology.

ECMWF also highlighted the operational efficiency difference: AIFS ENS can generate forecasts far faster than the physics-based forecasting system, with much lower energy use. For customers, that is not only a science milestone. It is a signal that AI ensemble guidance is becoming part of the operational forecast product set, not only an experimental chart layer.

What this means for GribStream users

GribStream supports both the deterministic AIFS Oper and the ensemble AIFS Ensemble. That makes it easy to:

• Compare AIFS-ENS vs. IFS ENS for the same events.
• Run backtests on AI guidance using as-of time travel.
• Build ensemble percentiles and event probabilities.

If your workflow spans July 2025 or May 2026, keep those boundaries explicit. July 1, 2025 marks the operational AIFS ENS v1 launch, while May 12, 2026 marks the AIFS ENS v2 change inside the same GribStream dataset timeline.

Sources

• ECMWF news: https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ensemble-ai-forecasts-become-operational
• ECMWF implementation page for AIFS ENS v1: https://confluence.ecmwf.int/display/FCST/Implementation%2Bof%2BAIFS%2BENS%2Bv1
• AIFS v2 fields on GribStream: /blog/aifs-v2-parameters-now-on-gribstream