Planetary Atmospheric Data Assimilation System

1
Planetary Atmospheric Data Assimilation System

• Mark I. Richardson and W. Gregory Lawson
• California Institute of Technology

2
Want Desire
Want

• Want to understand how planetary atmospheres
  work
• Dynamics of atmospheres
• Role of atmospheric motions in planetary climate
  systems

Desire

• Gridded accurate, frequent, dense data
• Certainly not directly available for planetary
  science
• Drives us to want to extrapolate the very limited
  available data

3
Challenge of using planetary atmospheric
observations

• Very sparse
• Remote sensing data (no current surface stations)
• Asynchronous
• Poorly known error characteristics (biases and
  (co-)variances)

4

• MGS 1999-2006
• MRO 2006-now

5
How best to extrapolate, spatially and
temporally?

• Binning and averaging - not really an
  extrapolation - just trading time and space
• Functional fitting - can extrapolate in an
  unphysical way
• Data assimilation - done properly, ought to
  physically (i.e. consistent with physical laws)
  extrapolate observations

6
What our project is

• To develop a ensemble-based data assimilation
  system for planetary use
• To optimize the system for the data-limited
  nature of planetary science
• To make the system easily adaptable for new
  planetary data sets and diverse atmospheric
  models
• To create a publicly available system amenable to
  use by (the typically small) planetary science
  groups
• The goal is not forecasting - were after
  reanalysis data for study of the atmosphere

7
The world of atmospheric DA
very, very high dimensional
discretized
high dimensional
images from Ross Bannister
8
Get confidence and run primer experiments with
simple model
Lorenz 96
advection
damping
forcing
dynamics
9
Ensemble Dispersion -- L96 is Chaotic! (or at
least sensitive to initial conditions)
Value (units)
180 W
180 E
0
Courtesy of DART team
10
Probabilistic
Monte Carlo
11
Ensemble-based Kalman filtering
time
12
Example of Data Assimilation using station
observations
Boston
Barcelona
Beijing
Value (units)
180 W
180 E
0
Courtesy of DART team
13
Ensemble approaches

• Pros
• Estimates state-dependent uncertainty
• Direct use of nonlinear operators (no TLM or AM)
• Theoretically pleasing -- slight non-normality
• Readily parallelizable competitive with other
  state of the art methods
• Modular / generalizable framework
• Poetry
• Monte Carlo methods are subject to sampling
  errors with which one must deal

14
Slides from J.L. Anderson
15
Slides from J.L. Anderson
16
Slides from J.L. Anderson
17
Slides from J.L. Anderson
18
Slides from J.L. Anderson
19
Slides from J.L. Anderson
20
What we are specifically doing

• Use a modular data assimilation tool from the
  National Center for Atmospheric Research (NCAR)
  the Data Assimilation Research Testbed (DART)
• Put a tested planetary model into DART
  (planetWRF) - the NCAR-based planetary Weather
  Research and Forecasting (WRF) model
• Introduce an example planetary dataset (the MGS
  Thermal Emission Spectrometer forward operator)

21
NCARs DART

• Explicitly embraces the modular, generalizable
  nature of ensemble DA, thus providing a framework
  within which to do DA

• DART Compliance
• Model must provide specific Fortran 90 interfaces
• Model must be callable as subroutine or drivable
  via shell script
• Forward Operator must be callable as subroutine
• Observations must be stored in certain format for
  Observation Sequence file

22
NCARs DART

• Offers many bells whistles
• Adaptive covariance inflation
• Tools for estimating proper localization
• An ensemble smoother (!)
• Maintained and supported by NCAR

23
planetWRF

• Example model - the DA system does not depend on
  WRF
• Grid-point, finite difference model
• Adapted for global, planetary (Mars, Venus,
  Titan) use
• Publicly available via NCAR

24
Thermal Emission Spectrometer
25
Where are we right now?

• Incorporated a terrestrial version of WRF in DART
• Tested DART for idealized systems and the
  terrestrial version of our WRF model
• Have obtained the TES forward operator and
  starting to strip it down for DART
• Have the latest version of planetWRF in testing
  mode for incorporation in DART

26
extra slides
27
Ensembles estimate state dependent uncertainty
28
Slides from J.L. Anderson
29
Sampling Error from Gaussianity

• Often addressed by Covariance Inflation

Slides from J.L. Anderson
30
Sampling Error from Linear Regression

• Often addressed by Covariance Localization

Slides from T. Hamill