Combining data assimilation and machine learning to estimate parameters of a convective‐scale model

Errors in the representation of clouds convection-permitting numerical weather prediction models can be introduced by different sources. These forcing and boundary conditions, orography, accuracy schemes determining evolution humidity temperature, but large contributions are due to parametrization microphysics processes surface layers. typically contain several tunable parameters that either not physical or only crudely known, leading model errors. Traditionally, values these chosen manual tuning. More objectively, they estimated from observations augmented state approach during data assimilation. Alternatively, this work, we look at problem parameter estimation through an artificial intelligence lens training two types neural network (ANN) estimate one-dimensional modified shallow-water as a function analysis atmospheric state. Through perfect experiments show Bayesian networks (BNNs) approximations point (NNs) able their relevant statistics. The combined with assimilation for decreases initial errors even when assimilating sparse noisy observations. sensitivity number ensemble members, observation coverage size is shown. Additionally, use method layer-wise relevance propagation gain insight into how ANNs learning discover naturally select few grid points subject strong winds rain make predictions parameters.