APPLICATION OF NEURAL NETWORKS TO EMULATION OF RADIATION PARAMETERIZATIONS IN GENERAL CIRCULATION MODELS

Abstract

A novel approach based on using neural network (NN) techniques for approximation of physical components of complex environmental systems has been applied and further developed in this dissertation. A new type of a numerical model, a complex hybrid environmental model, based on a combination of deterministic and statistical learning model components, has been explored. Conceptual and practical aspects of developing hybrid models have been formalized as a methodology for applications to climate modeling and numerical weather prediction. The approach uses NN as a machine or statistical learning technique to develop highly accurate and fast emulations for model physics components/parameterizations. The NN emulations of the most time consuming model physics components, short and long wave radiation (LWR and SWR) parameterizations have been combined with the remaining deterministic components of a general circulation model (GCM) to constitute a hybrid GCM (HGCM). The parallel GCM and HGCM simulations produce very similar results but HGCM is significantly faster. The high accuracy, which is of a paramount importance for the approach, and a speed-up of model calculations when using NN emulations, open the opportunity for model improvement. It includes using extended NN ensembles and/or more frequent calculations of full model radiation resulting in an improvement of radiation-cloud interaction, a better consistency with model dynamics and other model physics components. First, the approach was successfully applied to a moderate resolution (T42L26) uncoupled NCAR Community Atmospheric Model driven by climatological SST for a decadal climate simulation mode. Then it has been further developed and subsequently implemented into a coupled GCM, the NCEP Climate Forecast System with significantly higher resolution (T126L64) and time dependent CO2 and tested for decadal climate simulations, seasonal prediction, and short- to medium term forecasts. The developed highly accurate NN emulations of radiation parameterizations are on average one to two orders of magnitude faster than the original radiation parameterizations. The NN approach was extended by introduction of NN ensembles and a compound parameterization with quality control of larger errors. Applicability of other statistical learning techniques, such as approximate nearest neighbor approximation and random trees, to emulation of model physics has also been explored