There were five separate topic-oriented breakout groups in the User Voice Corner session at UEF 2026. Notes of the exchanges that took place are provided below, under the five topic headings.
Q: What is the difference between AIFS-Single and AIFS-ENS Control member?
Single is unperturbed, based on minimising an RMSE metric. AIFS-ENS Control is based on a different error (CRPS) minimization metric, and an unperturbed analysis, but has random "diffusion" perturbations applied during the run; it is therefore much less like a typical Control run, in the classical sense.
Q: Precipitation has improved with this cycle because now convective precip is moving. There are still some unrealistic bullseye blobs; is it possible that they will now move?
Yes blobs may still appear, e.g. over Finland. That aspect should have been improved a bit in 50r1, but there is a change that will do more in 51r1.
And yes, the bullseye blobs may now move.
Q. I loved the explanation for the difference between AIFS Single and AIFS ENS Control. It is difficult to be in a position where all these new models are published, and we struggle using all of this in our operational service. But we have to take a leap of faith on when to start using it in our operations. We are struggling to describe to our colleagues how to use it, when to use it, what the advantages of it are and when not to use it. Forecasters are just pushing products, we don't always have time to look at everything. Often we look at local model outputs and compare to IFS. The AIFS is just another (global) model. My question is, how do we get to learn it as forecasters. What's the main focus points we need to be aware of as forecasters.
Yes the main point is that all these models are very new. In addition you don't have the hindcast yet nor the maturity of IFS and experience of working with it. Currently, the main thing is to monitor and to learn how they perform. The temporal and horizontal resolution is low, and it will get better in the future.
Input from users would be good - so not all the forecasters have to learn individually (maybe ECMWF could organise a session on this).
AI models are quite good for 2m temperature, but not so great for 10m winds. For TCs they are good for the location, but not so great for winds.
Known AIFS model issues (and indeed known IFS model issues) are recorded and updated for users in the ECMWF Forecast User Portal here - see links in the top right.
Q: Do you expect that AIFS catches things earlier?
Sometimes the intensity of features (e.g. TCs) can be underestimated. Most of what it learns is from ERA5, and that uses a "low resolution" model (31km; whereas today's IFS runs at 9km).
Q: Can we expect more release of AIFS asyncronously than IFS. Will new release of AIFS come before IFS?
Yes, for example with 50r1 we had to release them together. The old AIFS was not coping well. This is something we still need to figure out. But we don't necessarily have to sync them. Let's see.
Q: Could you also take requests and feedback from the users. Now you have a new wave model; obviously from the moment people start to use it, the question of consistency will arise.
Of course user feedback influences our developments, and we also depend on it to some extent. We would all like higher temporal and spatial resolution.
Q: Regarding consistency among elements in the AIFS. When you're adding new elements in the AIFS, do you prefer to have consistency between them or to optimise each separately?
This is also an open science question. Currently we train AIFS-Single on RMSE for all the variables. We have a relatively simple cost function, so inconsistencies can arise. Indeed some have arisen. But at the same time the consistency achieved is usually quite high and quite reassuring. AIFS-ENS is trained in a different way, but inconsistencies can and have arisen there too.
Q: Is the IFS to AIFS nudging solution the answer to this or is ECMWF exploring other approaches to make AIFS and IFS more consistent?
We have introduced bounds on output values, to prevent negative precipitation, and to make convective and large scale precipitation consistent with the total precipitation. We are still working on future model configuration options, trying to decide upon the most expedient route, taking into account multiple different angles, including user needs, maintainability, computational tractability.
Q: Do you plan to make IFS to AIFS nudging model operational.
The decision has not yet been made. There are other hybrid approaches that we are exploring. For example where you look at differences between short-range forecasts and obs, and try to teach neural networks to learn model biases, and thereby to add other terms, to represent e.g. "missing" physics. At the moment the nudging approach looks much better in terms of scores.
One significant challenge is that the number of possible approaches is escalating. We have to manage this carefully.
Q: There are special cases with dust over Bulgaria, when IFS doesn't predict convection well, but for some reason AIFS does. In these cases dust influences the convective precipitaiton. I don't know if we have enough cases to say yes to one model, AIFS is better.
It might be worth looking at our CAMS forecast, and aerosols in that, though admittedly that may become too much information.
Users were asked to share those situations where dust appears to have been a key factor .
AIFS doesn't know anything about dust, but it can potentially learn behaviours indirectly.
Q: Is there now more advection of convective precipitation?
Some of the convective precip is now handed over to the large scale scheme, which will advect it. In the webinar on 50r1, there are some cases that were shown where you see the improvements, but the problem has not been completely solved.
Q: I am wondering how this will affect the west african cost? Is there anything that hints to the user that rainfall there will now be moving more (via advection).
Look for cases where there are strong winds over the coast. We haven't looked much at situations with light winds.
Q: Do the convective scheme change mean that 50r1 scatters precipitation over a wider area? Is the intensity is the same? Does precipitation with the same intensity expand to cover a bigger area?
Precipitation will cover a larger area. Overall intensities will be lower than previously, in some integrated sense (we are not creating extra rainfall). But of course inland (say) where in the previous cycle there may have been no rain now there will be some. So it becomes inevitable that the frequency of light precipitation/small precipitation totals goes up in cycle 50r1.
Q: Addition of 06/18UTC runs would be more useful if they had data beyond day 6.
This is somewhat constrained by ECMWF's computational resources.
Q: What is the difference between the number of people working on AI models versus the number working on physics-based models.
We have more people working on the physics-based models, but some people are working on both. And people can flip from one side to the other.
Q: What are the next big steps for AIFS?
Resolution improvements: time resolution first and then spatial resolution. And an ocean component in some form.
Q: What about improving convective precipitation in AIFS?
When you look at the extremes, it struggles. It is not predictable. The AIFS-ENS should not struggle, but most of the learning is on ERA5, which is at lower resolution. But this does not mean that we will not have something better in the future.
The new (ERA6) reanalysis will help; that has 14km resolution. It won't help with large scale aspects, but will bring in some higher resolution features. Although we augment ERA5 training with data from more recent 9km operational runs (10 years worth), it will be very helpful to have multiple decades of higher resolution data from ERA6.
Q: Will AIFS move to 9km resolution?
This is not planned for the next release of AIFS.
Q: CAT (Clear Air Turbulence parameters) have been quite well received, and it would be good to have them available in the AIFS.
Noted. To help with this we could learn not just from analyses but also from short range forecasts (which we already do for precipitation).
A representative from the World Energy and Meteorology Council (UEA, Norwich) raised the issue of extreme cold events, especially how well these can be predicted when they are linked to broader weather types, and asked if there was any verification for weather regime forecasts.*1,*2,*3
The representative from Serbia was working on regional climate change and reported that they have started monitoring forecast extremes (T2m, precipitation) using forecasts (e.g. EFI) as part of a rapid weather attribution study. Croatia highlighted an extreme wind event in northern Croatia that was well predicted seven days in advance. DMI (Denmark) identified an early cold spell in late October last year as a significant event and noted that the IFS appeared to perform well in that case, although capturing extreme low temperatures was often problematic in the forecasts – they do postprocessing, but it doesn’t do much for the extremes.
The representative from Croatia (an aviation forecaster) asked about the use of EFI wind speed for aviation forecasting, including how well the product performs and whether it has been verified. We explained that ECMWF regularly publishes an annual report that includes EFI verification results*4. We said the verification of EFI of gusts is more challenging.
The same representative noted that the CAT products are highly appreciated by forecasters. They also asked about an orographic gravity wave product, and we explained that the CAT product already contains a turbulence-mixing term related to orographic gravity waves*5. They were also interested in tropopause folding as a source of turbulence and that a 2D product identifying the location of tropopause folds could be useful; we noted that the 2 PVU product is available, but it does not explicitly identify tropopause folds, only the highest altitude of 2 PVU*6. The CAT product should already give higher turbulence estimates in these regions, but this could be explored further.
They also mentioned that aircraft icing products (not currently available from the IFS) would be very useful. The visibility probability meteograms were very much appreciated. The representative from Croatia also asked about forest fires, which are critical for airport operations if there are fires nearby. We showed the EFI fire product (eccharts), the Copernicus web (EFFIS), and CAMS as relevant resources.
Several representatives highlighted specific weather phenomena that remain particularly challenging in operations. For Montenegro, the main difficulties are fog, wind, ocean-waves (which have poor skill in the IFS in that part of the Adriatic). For the sub-seasonal forecasts cold episodes seem to be less well predicted than heat waves. The representative also noted that ECMWF is their main source of data. Overall precipitation guidance performs better than their limited-area model, although the IFS tends to over-estimate precipitation over Montenegro, and wind guidance is less satisfactory. In response to a question from Linus about the impact of heat waves on holidaymakers, the representative explained that this is not a major issue because most tourism is concentrated in coastal regions, while heat waves occur more often inland. The representative from Nigeria said precipitation is the most important forecast quantity for them and the most challenging forecast situations are related to precipitation and wind gusts associated with convective storms (noting they’ve seen a lot of storms so far in 2026). We mentioned experimental AEW forecasts and although there hadn’t been much interest in this, he thought other colleagues from the region could be interested. The representative from Mercator (Norway) said the ocean-wave model tends to underpredict the wave height (same for the ENS), and off the coast of Norway it was noted that the extreme waves associated with strong storms were not always correctly placed or represented well in the forecasts.
* Reference Material relevant for queries raised by participants (& passed to them):
*1) How far in advance can we predict changes in large-scale flow leading to severe cold conditions over Europe?
Laura Ferranti Linus Magnusson Frédéric Vitart David S. Richardson
Q J R Meteorol Soc.
2018;144:1788–1802. https://doi.org/10.1002/qj.3341
*2) Euro-Atlantic Weather Regimes and Their Modulation by Tropospheric and Stratospheric Teleconnection Pathways in ECMWF Reforecasts
CHRISTOPHER D. ROBERTS, MAGDALENA A. BALMASEDA, LAURA FERRANTI, AND FREDERIC VITART
*3) Year-round sub-seasonal forecast skill for Atlantic–European weather regimes
Dominik Büeler1 Laura Ferranti2 Linus Magnusson2, Julian F.Quinting1 Christian M. Grams1,
Q J R Meteorol Soc 147(741), 4283–4309.
Available from: https://doi.org/10.1002/qj.4178
*4) https://www.ecmwf.int/sites/default/files/elibrary/092025/81680-evaluation-of-ecmwf-forecasts.pdf
*5) Evaluation and Improvement of the ECMWF Aviation Turbulence Forecasts
Han‐Chang Ko1, Hye‐Yeong Chun1, and Peter Bechtold2
JGR: Atmospheres, 130.
https://doi.org/10.1029/2024JD043158
*6) https://confluence.ecmwf.int/display/FUG/Section+8.1.11+Potential+vorticity+charts
Overview:
The meeting focused on gathering feedback on ECMWF products, with particular emphasis on AI-based models, visualisation challenges, and user workflow efficiency. Participants discussed future directions for product development and how to better support operational forecasters.
1. AI models and explainability
- Strong interest in AI-based forecasting products (AIFS), but participants highlighted the need for explainability.
- Trust in forecasts depends on understanding why a model produces a given result, not just the output itself.
- There is currently a gap compared to traditional physics-based models where forecasters have well-developed conceptual understanding.
2. Data overload and efficiency
- A major concern is the increasing volume of data from multiple models, ensembles, and cycles.
- Forecasters do not have more time, so efficiency in accessing and interpreting data is critical.
- There is a need for smarter tools that filter, prioritise, and summarise information rather than adding more raw data.
3. Visualisation improvements
- Current workflows involve mentally integrating multiple datasets, which is inefficient and error-prone.
- Users requested more integrated views combining multiple parameters and models in a single interface.
- Better support for interactive exploration (zooming, layering, custom views) was highlighted.
4. AI for decision support
- AI could help guide forecasters by identifying critical regions, events, and relevant products.
- Concept of chatbot-style interfaces or query-driven dashboards to retrieve tailored information was discussed.
- Potential to automate selection of thresholds, projections, and relevant diagnostics.
5. Product strategy
- Important not to duplicate existing products in AI form without clear added value.
- Focus should be on novel diagnostics and summarised products (e.g. EFI-like indicators, event-based views).
- Products should serve broad needs first, with customisation layered on top.
6. Verification and trust
- Users are interested in real-time or near-real-time indicators of forecast skill.
- Desire for clearer linkage between forecasts and observations to assess reliability quickly.
- Objective verification and comparison between models were suggested enhancements.
7. Customisation and flexibility
- Demand for on-demand computation (e.g. clustering, regional focus, aggregation over different
time windows).
- Need for flexible meteograms and probability-based visualisations tailored to user thresholds.
8. Infrastructure and data access
- Preference for service-based access (querying ECMWF systems) rather than downloading large datasets.
- Importance of standard formats and interoperability across systems.
9. Resolution and regional needs
- Requests for higher resolution visualisation and access to native grid data.
- Regional customisation (e.g. Central Asia) identified as important for some users.
Conclusion:
The discussion highlighted a shift from data provision towards intelligent, user-centric products. Key priorities include improving explainability of AI models, reducing data overload through better summarisation and guidance, and developing flexible tools that align with real forecasting workflows.
Introductions & Roles:
Model Performance & Issues:
Agricultural Applications:
Extreme Temperatures & Diurnal Cycle:
Apparent Temperature & Regional Adaptation:
Model Development & Future Plans:
Verification & Observations:
Collaboration & Projects:
Summary:
Action Items from the Meeting