Title: Development of Neural Network Emulations of Model Physics Components for Improving the Computational Performance of the Ncep Seasonal Climate Forecasts

Consistently with the CTB goal “to accelerate the transition of research and development into improved operational climate forecasts, products and applications”, the proposed collaborative research will lead to improving the computational performance for operational seasonal climate predictions. The proposed study will take an advantage of the NCEP CTB that “will provide an operational testing environment to support ... applied research and development projects that will result in a direct influence on NOAA climate forecast operations, to be carried out jointly by scientists from operational centers and the broader research community.” This innovative research will be devoted to developing accurate and fast emulation of model physics using statistical learning techniques, namely neural networks (NN). It will be aimed at becoming an integral part of the operational NCEP CTB climate forecasting system. The proposed research will be based on the already existing and ongoing mutual collaboration between the NCEP and University of Maryland participants and their respective groups. The proposed collaborative research is considered by the proposers as a strategic scientific and methodological study with immediate practical applications to the NCEP operational system. Research highlights for the report period: • The current NCEP CFS model (GFS/MOM3) with NCEP’s versions of RRTM1 LWR and RRTM2 SWR (modified from AER Inc.’s RRTMG LW and SW radiation codes) have been used • Creation of NN training data sets for both LWR and SWR has been developed and adjusted for the coupled NCEP CFS model • The NN methodology and experimentation framework have been developed and refined • Refined NN emulations of both LWR and SWR have been developed and their accuracy estimated against the original LWR and SWR • A new measure of similarity for climate simulations and seasonal predictions, i.e., comparison with the interval model variability, has been introduced for comprehensive validation • Refined NN emulations for full model radiation (i.e. using both LWR NN and SWR NN) have been validated including comparison against the interval model variability; it was done for the parallel control (using the original LWR and SWR) and NN runs for the 17year (1990-2006) climate simulation and seasonal predictions