Institute for Mathematics Applied to Geoscience Geophysical Statistics Project - GSP Data Assimilation Research Section - DAReS Turbulence Numerics Team - TNT

Your typical ensemble

Ensemble mean = waste of time

GSP - statistically blending winds QuikSCAT Launched June 1, 1999 ~800 km altitude (~100 min orbit) 1800 km-wide swath about 400,000 measurements/day graphics by JPL

GSP - Statistically Blending Winds 6144 ‘observations’ every 6 hours (unrealistically smooth) ~ 40,000 observations every 6 hours (dense, but ‘gappy’) Resulting wind field with proper energy decay Gradient (important to ocean circulation) energy falls off too much energy falls off perhaps not enough

GSP The striping indicates the areas sampled by the scatterometer. Animation of E-W wind field created by synthesizing gridded winds and scatterometer (satellite) winds. Technique remains faithful to the data. Low standard deviation in data-dense regions.

Data Assimilation Research Section - DAReS Our computational challenge is to run MANY (~100) instances of the numerical models (CAM, WRF,...) simultaneously. Simply running one numerical weather prediction model has been driving supercomputer research. Data assimilation exploits the information in observations to ‘steer’ a numerical model. Put another way, it ‘confronts’ a numerical model with observations.

Our challenge is to run MANY of these.

Our ‘MACHINE’

Data Assimilation Research Testbed : DART * Many low-order models: Lorenz 63, L84, L96, etc. * Global 2-level PE model (from NOAA/CDC) * NCAR’s CAM 2.0 & 3.0 (global spectral model) * NCAR’s WRF (regional) * GFDL FMS B-Grid GCM (global grid point model) Forward Operators and Datasets Many linear, non-linear forward operators for low-models U, V, T, Ps, Q, for realistic models Radar reflectivity, GPS refractivity for realistic models Observations from BUFR files (NCEP reanalysis flavor) Can create synthetic (i.e perfect model) observations for all

Data Assimilation Research Section - DAReS

Turbulence is one of the last unsolved classical physics problems. GASpAR Geophysical-Astorphysical spectral element adpative refinement.

GASpAR Flexible framework for accurate simulation of turbulence Numerical methods that minimize dissipation. Objects are structured to facilitate parallel computation. Dynamic refinement gives a speedup of 5-10X over fixed grids with comparable accuracy. Objected-oriented h-adapted code for simulating turbulent flows.

Hierarchical. How does GASpAR do it? ElementsFieldsEquation Solvers Spectral Element Method operators GBLASBases Mortar objects Adaptive refinement based on error estimates

Mortar Objects An unambiguous representation of the field at parent/child boundaries based on interpolation.

Dynamically adaptive geophysical fluid dynamics simulation using GASpAR Simulation of three vortices Refinement done on a component of velocity.