RAL scientists continue to support the DoD’s National Ground Intelligence Center (NGIC) in its mission of assessing the consequences of the transport and dispersion of accidental and intentional releases of hazardous materials into the atmosphere. This is done by providing the agency with access to the RAL-developed GCAT (Global Climate Analysis Toolkit) system. GCAT is a fully automated dynamical downscaling system that allows NGIC scientists to remotely generate a high-resolution 30-year climatography for any region on the Earth. GCAT is based upon Climate Four-Dimensional Data Assimilation (CFDDA) technologies and can run with four domains, reaching a grid increment of 1.1 km. This enables NGIC to conduct transport-and-dispersion analyses at very fine scale. GCAT has the capability to automatically classify WRF output fields into climatological regimes. The method is based on the self-organizing map (SOM) [1] artificial neural-network pattern-recognition technique. Figure 1 shows the results of a SOM classification, in which 30 months (May 1981-2010) of WRF 1.1-km hourly output was used to estimate the main six regimes of the wind flow over the Energetic Materials Research and Testing Center in Socorro, NM. The six regimes that have been identified are given with their frequency of occurrence and their most representative days, which are chosen based on their Euclidian distance to each SOM node. Weather data valid for the representative days provides better forcing to NGIC’s transport and dispersion climatological studies because the data didn’t undergo averaging, which can destroy important model physical properties (balance etc.) available with dynamical downscaling.
The Second-order Closure Integrated PUFF (SCIPUFF) transport-and-dispersion model is implemented for execution for each dynamical downscaling simulation upon user request. This way, SOMs can be built based on plume dosage, in addition to weather variables, when analyzing the past climate. The system makes use of the Climate Forecast System Reanalysis (CFSR) data set available on a 0.5-degree grid for initial and lateral boundary conditions.
- Finalize the predictor weighting based on developmental testing described above;
- Couple the AnEn scripts with GCAT and offer it as a standard user-selected option;
- Test the updated GCAT climo system in a realistic HPC setting, which will include comparing with the 4-grid climo benchmark for accuracy and speed at various grid configuration settings; and
- Test the system out for a couple of other domains to examine its performance in other types of topography and climatological conditions.
As a more experimental part of development in the realm of computationally inexpensive downscaling, NCAR will examine as an alternative to the AnEn a machine learning approach to downscaling. In particular, the convolutional neural network (CNN) is well suited for this task since it can be used to identify important spatial patterns from the training data that relate to the patterns seen in the testing period. In other words, instead of conducting a grid-point based downscaling with the AnEn, the CNN would consider the entire domain in generating a downscaled, fine-scale meteorological field from coarse WRF grids. The training dataset (5 years of data for a month of interest) are sufficiently large for machine learning techniques. Moreover, the CNN may be able to provide an alternative to the SOM itself in defining typical days. The CNN has proven to be a powerful classification and diagnostic tool in various geosciences applications, including estimating hurricane intensity from satellite imagery, automatically classifying convective storm modes (e.g., supercell, bow echo, squall line) from radar imagery, and accurately estimating the probability of hail from radar imagery. It is now being used as a post-processing tool for numerical weather prediction as well.
- Implementing the seasonal ensemble forecasting system on Centennial.
- Testing with 1.5-month (sub-seasonal), 3-month (seasonal), and 6-month (2-seasonal) forecasting configurations.
- Evaluating the sub-seasonal, seasonal, and 2-seasonal ensemble forecasts using the analyses and observations.
- Issuing the sub-seasonal, seasonal, and 2-seasonal ensemble forecasts in near real-time.
- Testing the typical days' SOM classification based on seasonal forecasting. WRF simulations will use CSF forecast fields as initial and boundary conditions to generate downscaled predictions for a month in the future at different lead times (sub-seasonal, seasonal, 2-seasonal). The results will be compared to those obtained by the “standard” GCAT approach based on reanalysis of the last 30 years. We will focus this study on Dugway Proving Ground, where several weather stations are available.
[1] Kohonen T (1995) Self-organizing maps. Springer-Verlag, Heidelberg
[2] Jun Yan (2010). som: Self-Organizing Map. R package version 0.3-5. http://CRAN.R-project.org/package=som
[3] Alessandrini S, Hacker J, Vandenberghe F Definition of typical-day dispersion patterns as a consequence of a hazardous release, Proceedings of 34th International Technical Meeting on Air Pollution Modelling, Montpellier, France, May 2015.\
[4] Ferrero, E., Vandenberghe, F., Alessandrini, S. and Mortarini, L., 2016, December. Comparison of WRF PBL Models in Low-Wind Speed Conditions Against Measured Data. In International Technical Meeting on Air Pollution Modelling and its Application (pp. 129-134). Springer, Cham.