Classical numerical forecast models (e.g. IFS) assume there is a reasonably accurate physical model of the Earth system. The biggest unknown is the initial conditions from which to start the forecast.

Within Machine Learning (ML) forecast models (e.g. AIFS) the physics of the atmosphere is not explicitly handled. The aim is to continually develop (“train”) an empirical model directly from the observations (or a surrogate for those, such as reanalyses).  Whilst observations implicitly contain the physics of the atmosphere, it is not necessary for ML models to "understand" the underpinning physics that dictates the evolution of variables through a forecast.  Nevertheless, ML models do seem to be learning some physics as they develop.  In effect ML models use previous evolutions of similar analyses at several levels to “learn” the more likely evolution.

Making a forecast with ML models is very efficient. They require only a single Graphics Processing Unit (GPU), take less than a minute to run, and consume a tiny fraction of the energy required for an IFS forecast.

However, physics-based numerical weather prediction models are still key for these fully ML approaches.  The IFS is used to create both training and validation data (predominantly ERA5).  AIFS and other machine learning models were trained using ERA5 and a few years of operational analysis to minimise the mean squared error of forecast parameters.  In this way they have been trained for use as a deterministic model.  Moreover, after the training process, ML models rely on the IFS to provide initial conditions for each forecast.

Physics-based numerical weather prediction models are still key.  The IFS is unparalleled for the breadth of variables it predicts and for its spatial resolution.  ECMWF remains firmly committed to further improvement of the IFS.

The ECMWF AIFS model uses Graph Neural Networks allowing grid-flexibility and parameter efficiency.  Current ML models run much faster and much more cheaply than traditional physics based forecast models.  However, currently they have lower temporal, horizontal and vertical resolution than the IFS models.  Multi-date verification suggests ML broadscale forecasts score better than classical NWP but at the expense of shorter wave features and fine detail as forecast lead time increases.

ML models are evolving quickly.  Users should try to stay up to date with the latest AIFS developments.

"Deterministic" ML model

How does it work?

  • Training tries to minimise an error metric, considering all variables at the same time, with weightings applied to different variable types.  Most of the training focuses on T+6h, although later evolution, up to T+72, does also come into play at the very end of the training.
  • The parameters used in training include pressure-level data for basic variables plus a few extra surface variables at every ERA5 grid-box, every 6 hours.
  • Forecast production is based on using interrelationships between physical variables in consecutive time windows that were learnt during the training period.
  • Currently all the manpower that went into creating the re-analysis used for training, (i.e. tens of thousands of person years work worldwide), underpins all that an ML model can do.  Performing some form of data assimilation by ML methods could change that in the longer term).

Current Characteristics of AIFS

  • High quality broadscale forecasts that beat classical NWP over multi-date verification (e.g. better handling of large scale wave propagation).
  • Successive forecasts tend to be less jumpy than classical NWP.
  • Very fast and economical to produce a forecast sequence (though training takes day/weeks/vast computer power).
  • Resolution is innately very low; pressure levels in vertical, ~ERA5 in horizontal (0.25 deg).  This does not innately reduce the skill of the output, compared to what would happen in classical NWP if there were only a few pressure levels.  But it does reduce the capacity to deliver certain products to a useable standard (e.g. vertical profiles).
  • Fine-scale details that one would see in HRES will be missing (e.g. for precipitation).
  • Many desirable parameters are currently missing or are unreliable (e.g. MUCAPE, convection, convective precipitation, cloud, visibility, precipitation type, snow, gusts etc.).  However, this may well be a short-term limitation.
  • At longer lead times:
    • Increasingly smooth look to fields - increasingly 'implausible' output (even if RMSE looks very good).
    • Fewer, smoother short wavelength features.
    • Reduced gradient strengths (e.g. frontal zones systematically broaden then disappear).
    • Some extreme weather events appear less intense (e.g. windstorms which appear less intense than IFS but are nevertheless indicated).  However, temperature extremes are more consistently handled, generally beating classical NWP deterministic forecasts.

Practical use of AIFS output

Forecasters should:

  • Focus on known strengths (e.g. broadscale pattern, tropical cyclone paths).
  • Steer well clear of more 'sophisticated'/detailed outputs (e.g. looking at rain over upslopes).
  • Try to link broadscale pattern to 'weather' or indeed other classical NWP output fields (e.g. in clusters/regimes that match).  In essence a type of fingerprinting approach might be useful.
  • Be very wary of using output in a direct way to predict most types of weather extremes.  2m temperature extremes are the exception - forecasts of these are often as good as IFS.
  • Be ready for unforeseen outcomes (e.g. extreme noise, semi-permanent upper low over Middle East, precipitation totals not adding up).
  • Expect the unexpected (e.g. odd behaviours, at least from time to time!).
  • Don't expect inter-variable correlations to always match up with classical ideas or classical NWP.  ERA5 learning should have reduced discrepancies but anomalies do occur (e.g. currently it is quite common to see convective precipitation greater than total precipitation.
  • Note that guidance could change from day to day.  ML models continue to develop and improve with corresponding changes in results. 

 

  • No labels