Correcting weather models by learning nudging tendencies from hindcast simulations

Abstract:

Coarse-resolution weather forecast models can have substantial biases in their simulated mean state. If these biases manifest themselves quickly enough, they can degrade short-term weather forecasts. We attempt to bias-correct the 200km FV3GFS atmospheric model using a hindcast simulation nudged towards the GFS analysis. The nudging acts on the temperature, specific humidity, surface pressure and horizontal wind fields. It tends to heat and dry the atmospheric column, particularly in regions of convection and extratropical fronts. A random forest is trained to predict column profiles of the temperature and humidity nudging tendencies from this simulation. A forecast is then made with the FV3GFS model coupled to the random forest predicting a heating and moistening correction at each time step. This ML-assisted forecast runs stably for at least a month, has only slight drifts in global mean temperature, and shows improved skill (compared to an equivalent forecast without the random forest) at predicting the large-scale patterns of atmospheric circulation on 5- to 10-day timescales. In addition, the diurnal cycle of precipitation over land is more realistic in the ML-assisted run.