"New ECMWF Model Versions: 50r1 and AIFS V2s"
Q1: 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.
...
Q3. 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.
A: 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.
...
Q: Can we expect more release of AIFS asyncronously than IFS. Will new release of AIFS come before IFS?
A: 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.
A: 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?
A: 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?
A: 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.
A: 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.
...
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.
...