Investigating the Value of Adding Diverse ML Models to the IMPROVER Blend

The central research question is: Does incorporating a broader and more diverse range of machine learning weather forecasting models into the IMPROVER blend deliver sufficient forecast improvement and operational value to justify the additional complexity of implementation and operation? The Met Office is particularly interested in understanding: • The value of model diversity within forecast blends. • Whether AI models provide additional skill when blended with existing NWP systems. • The marginal benefit of adding successive AI models. Whether a point of diminishing returns exists beyond which additional models provide little operational value. • Which combinations of models deliver the greatest forecast benefit. • Where benefits are largest? For example, at what lead times and for what parameters? • Whether complementary behaviour between models is more important than individual model skill alone. • Whether increasing the number of MLWP members in the blend could compensate for fewer NWP ensemble members in the blend The work should focus on operationally relevant questions rather than purely theoretical investigations. While forecast skill improvements are important, the ultimate goal is to determine whether these improvements are sufficient to influence future operational decisions. The outputs should therefore provide evidence that can support future Met Office strategy. Particular interest exists in forecasting lead times beyond approximately three to five days, where traditional NWP systems begin to lose skill and where MLWP models may provide the greatest opportunity for improvement. However, shorter lead times should not be excluded if evidence demonstrates that complementary model behaviour leads to measurable forecast improvements. All meteorological parameters are of interest and IMPROVER offers the opportunity to blend a different number of models for different parameters. The core variables of wind, temperature and precipitation are of most interest. We are interested in general 'every day' performance as well as extremes, particularly over the UK.