GribStream

GribStream Blog

ECMWF will stop experimental GraphCast and FourCastNet charts when AIFS v2 launches

|
aifsoper aifsenfo graphcast fourcastnetgfs

ECMWF's IFS 50r1 and AIFS v2 implementation pages say its experimental Aurora, FourCastNet, GraphCast, and Pangu-Weather chart models will stop when the May 12, 2026 upgrade goes live.

ECMWF's implementation pages for IFS Cycle 50r1, AIFS Single v2, and AIFS ENS v2 now say that, as part of the May 12, 2026 upgrade, ECMWF will stop running its experimental Machine Learning model charts:

  • Aurora
  • FourCastNet
  • GraphCast
  • Pangu-Weather

We already covered the May 12 cutover in ECMWF sets May 12, 2026 for IFS Cycle 50r1 and AIFS v2. The important new detail is the explicit shutdown of those ECMWF experimental ML chart products.

What is changing at ECMWF

The same May 12 upgrade moves ECMWF's operational AI line forward:

  • AIFS Single moves from v1.1 to v2
  • AIFS ENS moves from v1 to v2
  • both AIFS lines add a 10 hPa pressure level
  • both add a new wave component/stream
  • AIFS ENS v2 adds an operational data-driven wave ensemble stream

That means ECMWF is consolidating operational AI forecasting around AIFS Single and AIFS ENS, while the separate experimental chart models are being turned off.

Why ECMWF is stopping the external models

ECMWF has since published a more detailed AIFS blog post explaining the decision. The short version is that the IFS 50r1 upgrade changes the analysis used to initialize these models, and that created a domain-shift problem for some of the external machine-learning systems.

Before the cutover, ECMWF ran Pangu-Weather, GraphCast, Aurora, and FourCastNet in an experimental real-time setting, generally initialized from ECMWF operational analysis. ECMWF tested the models with pre-operational 50r1 analysis over January 21 through March 20, 2026 and compared the behavior with the 49r1-initialized runs. The result was not a simple win or loss across the board: Pangu-Weather and FourCastNet were less affected, while GraphCast, Aurora, and AIFS v1.1 showed degraded performance after the initialization change.

That detail matters for users comparing AI forecast families. ECMWF is not saying "ML weather models are over." It is saying first-generation external chart models tied to an older analysis environment are less useful once the operational IFS analysis changes. AIFS v2 was fine-tuned with prototype 50r1 data for exactly this reason, and ECMWF points to a broader operational AI ecosystem that now includes systems such as DWD AICON, Environment and Climate Change Canada's GEML, and NOAA's AIGFS/AIGEFS.

Does this discontinue GribStream GraphCast or FourCastNet?

Not by itself.

The ECMWF pages are talking about ECMWF's own experimental Machine Learning charts. GribStream's GraphCast and FourCastNetGFS pages refer to NOAA/NCEP GFS-based datasets, not those ECMWF chart experiments.

So this ECMWF notice does not announce that GribStream's GraphCastGFS or FourCastNetGFS data stop on May 12.

If the NOAA/NCEP GraphCastGFS or FourCastNetGFS datasets are discontinued separately, GribStream would stop receiving new forecasts for the discontinued feed. We would keep serving historical forecasts for as long as the underlying files remain available through the AWS Open Data program.

It still matters because the names overlap. If you compare GribStream AI datasets against ECMWF's experimental ML chart products, that comparison set may disappear. If you use ECMWF AI guidance operationally, the replacement path is AIFS Single v2 and AIFS ENS v2, not the old experimental chart set.

For GribStream users, the distinction is now clearer:

What to watch

  • The May 12, 2026 cutover for IFS 50r1, AIFS Single v2, and AIFS ENS v2
  • Any ECMWF follow-up on exact chart/product retirement behavior
  • Separate NOAA/NCEP or AWS Open Data notices for GraphCastGFS or FourCastNetGFS availability

For GribStream users, the safe assumption is: ECMWF's experimental ML charts are being retired, while NOAA/NCEP GraphCastGFS and FourCastNetGFS should be tracked through NOAA/AWS sources separately.

Sources

Related models

aifsoper aifsenfo graphcast fourcastnetgfs
Try GribStream