In this study, a three-layer NN was trained to reproduce the behavior of the Thompson Microphysics (TM) scheme in the WRF-ARW model (Thompson 2008). After determining the number of hidden neurons necessary for the statistical modeling of the TM scheme, we evaluated the performance of the NN-MP scheme through an off-line assessment of the variance explained in the behavior of the original TM scheme, and by a comparison of the WRF-ARW model skill using TM vs NN-MP.
If tested successfully, the NN-MP scheme can potentially be used to (a) speed up the performance of MP schemes in NWP models, (b) represent uncertainty related to subgrid-scale (stochastic) processes in ensemble forecasting, and (c) constrain the covariance of microphysical and other variables in variational data assimilation. In subsequent studies, the performance of NN-MP will also be verified against and improved by relevant observational data.
Acknowledgements: Steve Albers, Geoff Dimego, Jun Du, Brad Ferrier, Mark Govett, Isidora Jankov, Kevin Kelleher, and Gregory Thompson provided valuable input to and support for this project.
Krasnopolsky V., 2013. "The Application of Neural Networks in the Earth System Sciences. Neural Network Emulations for Complex Multidimensional Mappings", Springer.
Thompson, G., P. R. Field, R. M. Rasmussen, and W. D. Hall, 2008: Explicit Forecasts of Winter Precipitation Using an Improved Bulk Microphysics Scheme. Part II: Implementation of a New Snow Parameterization. MWR, 138, 5095-5115.