Machine Learning for emulating Physical Parameterization of Planetary Boundary Layer Height

Most models today take aerosol and chemical forcing into effect through physical parameterizations with many uncertainties in the parameters such as size and distribution globally. Resolving the uncertainties of the role of aerosols in current cloud resolving models at high resolution needing sub-grid scale requires the enormous computing resources. Machine Learning emulator of parameterizations has shown to be more computationally feasible. However, its ability to discover selected forecast features of variables by learning from the complex microphysics parameterizations still needs to be validated. We have developed a Neural Architecture Search (NAS) Machine Learning (ML) Model to emulate the Planetary Boundary Layer Height (PBLH) parameterization using outputs from the Weather Research and Forecasting model coupled with Chemistry (WRF-Chem) model. WRF-Chem uses the Mellor-Yamada and/or Yonsei University Planetary Boundary Layer scheme. The output from the WRF model has been run at a resolution of 0.8 degree over the US with lateral boundary conditions at a resolution of 0.25x0.25 degree from the NOAA GFS model. The NAS Regression Model uses Bayesian optimization to guide the network morphism for searching for the optimum neural architecture, which should reduce the computational parameterization computer time. At the end of training, the NAS search algorithm should give the best ML model emulation of the given input data. We use dense blocks for the neural computation, and one-way feed forward connections between layers. Potential temperature, winds in x, y, z directions of 29 vertical layers in the columns are used as input parameters for training the ML algorithm to predict the PBLH. Our NAS ML emulator results show a 40% to 45% error reduction, and a Pearson correlation improvement of 16% to 21% compared with Feed Forward Neural Network developed by the developers. The overall percent of NAS ML's PBLH prediction errors are ~5% compared with WRF Chem model output's PBLH.