Neural Network Applications for Numerical Weather Prediction V

Neural Network Applications for Numerical Weather Prediction V. Krasnopolsky

Abstract This presentation briefly overviews major AI/NN application developed and/or are being developed in three prime subfields of Numerical Weather Prediction (NWP): model initialization, numerical model, and model output post processing. Significant applications are listed and briefly discussed, several examples are presented. Advantages and limitations of AI in NWP fields are briefly discussed. 2

I. NN for Model Initialization • Developed NN Applications – Satellite Retrievals • SSMI retrieval algorithm (operational since 1998) • Quick. Scat retrieval algorithm – Direct Assimilation • Forward model for direct assimilation of SSMI BT • Quick. Scat forward model – Better Assimilation of Surface Observations • Observation operator for assimilation of SSH anomaly • Empirical biological model for ocean color • NN algorithm to fill gaps in ocean color fields and for creating long and consistent ocean color data sets April 25, 2019 V. Krasnopolsky. NN Applications for NWP Models 3

II. NN for Numerical Model • NN Applications Developed (black) and Under Development (red) – Fast and accurate emulations of parameterizations • Fast nonlinear wave-wave interaction for Wave. Watch • Fast NN long and short wave radiation for NCEP CFS, and GFS models and for FV 3 GFS • NN emulation super-parameterization (CRM in MMF) • Fast NN microphysics for FV 3 GFS and WRF – New parameterization • Convection parameterization for NCAR CAM learned by NN from CRM simulated data – NN emulation of simplified GCM April 25, 2019 V. Krasnopolsky. NN Applications for NWP Models 4

III. NN for Post-processing • NN Applications Developed (black) and Under Development (red) – Nonlinear ensembles • Nonlinear multi-model NN ensemble for predicting precipitation rates over Con. US • Nonlinear NN averaging of wave models ensemble • Nonlinear ensemble for hurricanes – Nonlinear bias corrections • Nonlinear bias corrections for GFS • Nonlinear approach to improve CFS week 3 an 4 forecast April 25, 2019 V. Krasnopolsky. NN Applications for NWP Models 5

What is common for aforementioned applications? They all are mappings: • Mapping: A rule of correspondence established between two vectors that associates each vector X of a vector space with a vector Y of another vector space AI tools: NNs, Support Vector Machines, Decision Trees, etc. are generic tools to approximate complex, nonlinear, multidimensional mappings. 6

Several Examples 1. NN Observation operator allows: a. Instantaneously propagate impact of surface observations vertically b. Assimilate bio and chemical observations that currently do not have corresponding prognostic variables in the model (create empirical chemical and biological models) 2. NN emulations of physical parameterizations allow to create fast model physics 3. NN ensemble averaging demonstrated impressive skills to improve ensemble predictions. April 25, 2019 V. Krasnopolsky. NN Applications for NWP Models 7

DAS: Propagating Information Vertically Using NNs, Assimilating Chemical and Bio data “Ground” Observatio ns (mainly 2 D) Satellite Data (2 D) N N 1 Model Predictions (3 D & 2 D) N N 2 Ocean DAS NN 1 and NN 2 – observation operators April 25, 2019 V. Krasnopolsky. NN Applications for NWP Models 8

Accurate and Fast NN Emulations of Model Physics • Any parameterization of model physics is a mapping and can be emulated by NN • The entire model is a mapping and can be emulated by NN • NN emulation is usually 1 to several orders of magnitude faster than the original parameterization: ~20 times for SWR and ~100 times for LWR April 25, 2019 V. Krasnopolsky. NN Applications for NWP Models 9

The Magic of NN Performance Xi Original Parameterizati on Input/Output Dependency: Yi Y = F(X) Xi NN Emulation Yi NN Emulation of Input/Output Dependency: YNN = FNN(X) Numerical Scheme for Solving Equations Mathematical Representation of Physical Processes Input/Output Dependency: April 25, 2019 V. Krasnopolsky. NN Applications for NWP Models {Xi, Yi}I = 1, . . N 10

LWR Individual HR Profiles Black – Original Parameterization Red – NN with 100 neurons Blue – NN with 150 neurons PRMSE = 0. 18 & 0. 10 K/day April 25, 2019 PRMSE = 0. 11 & 0. 06 K/day PRMSE = 0. 05 & 0. 04 K/day V. Krasnopolsky. NN Applications for NWP Models 11

The following figure: (a) 17 year run with the RRTMG LWR and SWR (b) 17 year run with the NN LWR and NN SWR (c) Run (a) - Ran (b) (d) Difference of 2 runs with the RRTMG LWR and SWR and different versions of FORTRAN compiler 12

CTL NN FR NCEP CFS PRATE – 17 years JJA NN - CTL April 25, 2019 V. Krasnopolsky. NN Applications for NWP Models CTL_O – CTL_N 13

Post-processing Model Output • Nonlinear multi-model ensembles for precipitation forecast. • Precipitation forecasts available from 8 operational models: – NCEP's mesoscale & global models (NAM & GFS) – the Canadian Meteorological Center regional & global models (CMC & CMCGLB) – global models from the Deutscher Wetterdienst (DWD) – the European Centre for Medium-Range Weather Forecasts (ECMWF) global model – the Japan Meteorological Agency (JMA) global model – the UK Met Office (UKMO) global model • Also NCEP Climate Prediction Center (CPC) precipitation analysis is available over Con. US. April 25, 2019 V. Krasnopolsky. NN Applications for NWP Models 14

Example of NN 24 hour forecast Verification CPC analysis NN ensemble April 25, 2019 Conservative ensemble Human analyst V. Krasnopolsky. NN Applications for NWP Models 15

Conclusions • NN is a generic and versatile AI technique • There exist numerous applications in Numerical Weather Prediction that can be and have been successfully approached using NNs • In NWP models NNs can be used in model initialization, as parts of the model physics, and for post-processing model outputs • A significant experience in developing NN applications for NWP models has been accumulated at NCEP April 25, 2019 V. Krasnopolsky. NN Applications for NWP Models 16

There is no free lunch • NN, as any statistical model, requires data for training; it is a data driven method • NN, as any nonlinear statistical model, requires more data, than linear model/regression • As any nonlinear statistical model, NN may be over fitted • As any statistical model, NN should be periodically updated to changes in environment; NN can be updated on-line April 25, 2019 V. Krasnopolsky. NN Applications for NWP Models 17