Machine learning based algorithms for uncertainty quantification in numerical weather prediction models

Complex numerical weather prediction models incorporate a variety of physical processes, each described by multiple alternative schemes with specific parameters. The selection the and choice corresponding parameters during model configuration can significantly impact accuracy forecasts. There is no combination that works best for all times, at locations, under conditions. It therefore considerable interest to understand interplay between physics resulting forecasts different This paper demonstrates use machine learning techniques study uncertainty in due interaction processes. first problem addressed herein estimation systematic errors output quantities future this information improve second considered identification those processes contribute most forecast quantity specified meteorological In order address these questions we employ two approaches, random forests artificial neural networks. discrepancies results observations past times are used learn relationships errors. Numerical experiments carried out Weather Research Forecasting (WRF) model. precipitation, variable both extremely important very challenging forecast. consideration include various micro-physics schemes, cumulus parameterizations, short wave, long wave radiation schemes. demonstrate strong potential approaches aid