Main contrasts between Classical and Artificial Intelligence (AI) forecasting models 

Traditional numerical weather forecasting relies upon a representation of the state of the atmosphere as accurate as possible at a given time.  Initial status of each model forecast is based upon observations and upon background fields of parameters derived from previous executions of the forecast suite.  This analysis is followed by time-step iterations using a simulation of physical laws and fluxes of heat, moisture and momentum.  The ECMWF Integrated Forecast System (IFS) model is a complex but refined representation of the interactions in the model atmosphere.  It is fully interactive with energy fluxes from land and sea surfaces.  Computations are repeated at small simulated time-steps at a large number of grid points across the globe.  Because of the complexity, it is relatively expensive in time and resources.  

Artificial Intelligence weather forecasting models are radically different to the physics based techniques currently widely used.  The aim is to define an empirical model that links observational data at one time with similar data for a short time later.  The ECMWF Artificial Intelligence Model (AIFS) has been developed (or trained) using reanalysis fields as observations and then compared with reanalysis fields for the later time.  Having found this empirical model it can be used iteratively to compute forecast fields at successive, fairly large, time steps into the future.  There is no need to explicitly simulate the physics of the atmosphere;  this is currently implicit in the reanalysis training data used.  The user may not even know the impact of the local or large scale physical effects upon the process.  Artificial Intelligence (AI) and Machine Learning (ML) are able to deal with massive amounts of data economically.  There are major speed and cost benefits.  Artificial Intelligence is destined to play a substantial role in meteorological forecasting in the future.  Research towards hybridising data-driven and physics-based forecasts will continue over the coming years.

There are advantages to both IFS and AIFS techniques.

IFS:

  • the high (9km) resolution can resolve relatively small meteorological features.
  • the physical basis and resolution allows finer detail for surface products.

AIFS:

  • gives better results in tropical cyclone tracking.  Accuracy can be higher than physics based models by up to 20%.
  • gives better deterministic and ensemble prediction of many standard upper-air parameters (e.g. geopotential height and wind).
  • is an extremely fast process and provides forecast output very rapidly.


Fig2-1: Root mean square error of forecast 2m temperatures (degC) over Europe during winter quarter December 2025 to February 2026.  IFS forecasts (red) and AIFS forecasts (blue).  AIFS shows a mean error about 0.3C less than IFS (abut 6% difference). But the diagram does not show the advantage of IFS in forecasting finer detail variations.  Even at day1 both IFS and AIFS show about 2.1C mean error.


Fig2-2: Northern hemisphere Anomaly Correlation Coefficient (ACC) for geopotential height at 500 hPa of IFS forecasts (red) and AIFS forecasts (blue) for 2022.  Higher values indicate better skill.


(FUG associated with Cy50r1)

  • No labels