Effect of Predicting Surface Humidity on ML Forecasts of Humid Heatwaves

Ascertain whether the observed performance degradation of machine-learning weather forecasting models in forecasting the 2023 South Asian humid heatwave persists when using ML models that predict surface-level humidity, specifically evaluating whether direct prediction of surface relative humidity mitigates underestimation of heat index values.

Background

Accurate assessment of humid heat hazards requires the heat index, which depends on 2-meter temperature and surface-level relative humidity. Current ML models evaluated (GraphCast and PanguWeather) do not predict surface humidity, necessitating approximations using 1000 hPa level humidity, which led to underestimation of heat index during the South Asian humid heatwave.

The unresolved question is whether ML models that directly output surface humidity would avoid this underestimation or whether extrapolation challenges would still degrade performance.