DOC

CURRICULUM VITAE

By Ronnie Patterson,2014-03-26 20:05
10 views 0
Steve introduced NCEP/EMC DA activity and plan. Various issues have been raised observations or simulated observations in a perfect model framework?

    SUMMARY OF TELECONFERENCE DISCUSSION

    December 18, 2008, 11:00am 1:00 pm EST

    Present: J. Whitaker and T. Hamill and X. Wang(NOAA ESRL), C. Bishop and Dan Hodyss (NRL), M. Zupanski (CIRA/CSU), I. Szunyogh and E. Kalnay(UMD), Z. Toth, M. Wei and M. Pena (NCEP/EMC)

Agenda:

    1) Review salient issues from Buenos Aires meeting (30-45 mins) 2) Review of new developments/results from each group since last meeting

     (10-15 mins max each)

    3) Discussion on coordination among groups

Eugenia Kalnay:

    Introduced the Workshop on 4D-Var and ensemble Kalman Filter Inter-comparisons which was held in Buenos Aires, Argentina from 10-13 November 2008. The outcome from this workshop demonstrated the comparative performance of EnKF method compared with variational method which is widely used at NWP centers. Most of the time was focused on the inter-comparison of 4D-Var and ENKF systems at Canadian system presented by Mark Buehner. Both 4D-Var and EnKF have been operational at CMC since 2005, both use the same forecast model and similar set of observations. Thus the two systems at CMC provide a quite clean comparison between the two methods. The conclusion is that 4D-Var and EnKF analyses have comparable quality, while 4D-Var with flow-dependent EnKF covariances yields a gain of about 10 hours at day 5 in southern extra-tropics. Work is still on the way to complete EnKF experiment using incremental approach to produce high-resolution deterministic analysis. The results from JMA show EnKF has worse results in SH, but better in NH due to the model bias problem. Overall, participants were excited to see how close the results produced by the two methods have become.

    Her work at University of Maryland includes the work on CO2 in DA with SPEED model. The goal is to compare LETKF with the inverse method.

Craig Bishop:

    Sent slides showing the comparison of localization in radiance space and model space. Explained why adding satellite radiances improve more in variational DA system than EnKF system. The results show the radiance space localization is inferior to model space localization.

     1

Jeff Whitaker:

    Spent more time on software engineering, as a result EnKF code is one-order of magnitude faster than last year. He is focusing more on bias correction and applying the EnKF system on hurricane study. Future work will include the development of hybrid approach. Pointed out that there are differences between the ways of ensembles being used at the CMC and NCEP. The CMC method take full advantage of flow-dependent information from the ensembles.

    Later on Dec. 30, 2008, Jeff sent slides that show great improvement of EnKF over GSI and the speed of codes is much faster than before. See the slides (on the same website) for more details.

Milija Zupanski:

    Introduced his recent work, focused more on much higher resolution of DA targeting for microphysics processes at 9-13km. At this resolution, EnKF produces much better picture than GSI. In his method, the LETKF is used in sub-domain smoothing.

Istvan Szunyogh:

    His work includes working on chemical DA with OZ, but pointed out there are challenges in localization for OZ. It will be interesting to study the vertical localization. Another area is EnKF in coupled global system that involves both atmosphere and ocean. He is also continuing his work on predictability. The effort is on using ensemble to represent the uncertainty in the system.

Zoltan Toth:

    Summarized the tele-conference and the activities in the past year. He congratulated Tom Hamill who will take over as the manager of NOAA Thorpex activity from next year. Suggested the next tele-conference in March 2009.

    Tom Hamill expressed his thanks to Zoltan Toth for his work during the past years in organizing NOAA Thorpex DA activities.

    Mozheng Wei’s slides on analysis differences and error variation from different NWP centers was postponed to the next time due to the time limit.

     2

    SUMMARY OF TELECONFERENCE DISCUSSION

    April 1, 2008, 1:00pm 3:00 pm EST

    Present: J. Whitaker and T. Hamill (CDC), C. Bishop (NRL), M. Zupanski (CIRA/CSU), I. Szunyogh and E. Kalnay(UMD), Y. Song, Z. Toth, M. Wei and Steve Lord (NCEP/EMC), E. Kalnay (UMD). Gregory Hakim (Uni. of Washington)

Agenda:

     1) Introduction of DA activity and plan by Steve Lord.

     2). Review of progress by each group (5-10 minute each group with slides

     if possible)

    2). Plans for the next 3-6 months by each group

    3). Longer term issues-expected benefits & computational costs of ens-DA

     and hybrid variational approach.

Steve Lord:

    Steve introduced NCEP/EMC DA activity and plan. Various issues have been raised and discussed. EMC has been developing next version of GSI called FOTO (First-Order Time extrapolation to the Observation). FOTO is a relatively efficient technique for taking into account temporal distribution of the observations within a 3D-VAR data assimilation system. The experimental results demonstrate the positive impact over the current operational GSI. FOTO is expected to be implemented in May 2008. EMC is continuing cooperation with NASA DA Office in development of tangent linear model of GSI towards an eventual full version of 4D-Var depending on its final performance. He also talked about the difficulty of contributing more computing resources to Thorpex funded ensemble DA activities before the arrival of the NECP new supercomputers which is expected to be in June 2008.

Milija Zupanski:

    Sent slides showing the application of localized MLEF plus WRF regional model to Hurricane Katrina. In this method, he used error covariance localization based on a modified local domains approach. The domain studied is divided into 9 local domains with each containing 23x25 horizontal grid points. The

     3

    observation weight matrix is a linear interpolation from neighboring local domains. Positive impact has been found in application to Hurricane Katrina case study.

Jeff Whitaker:

    New slides showed the comparison results between GSI, LETKF and EnSRF. Code has been ported to NCEP system and optimized. The comparison can be made with operational GSI and focus more on forecasts of hydrologic variables like humidity, precipitation and probabilistic forecast verification. The results are for period from May 15 to June 15, 2007. The model resolution is T126L64 with 64 ensembles. There is almost no difference between EDA and GSI in terms of forecast RMS errors verified against own analyses. For RMS error of precipitable water, EDA forecasts are significantly more skillful than GSI. EDA is much better in tropics and significantly better at short leads time over CONUS. In terms of ensemble forecasts of T850, EnKF skill is slightly larger at the shortest lead time but quickly loses to ET. ET spreads grow more quickly than EnKF. Overall, computational cost for EnKF has been reduced.

Eugenia Kalnay:

    Sent some slides and a short paper describing a new method for accelerating the spin-up of ensemble Kalman filter so that it even converges faster than 4D-Var. It uses a “running in place” algorithm which is made possible by the use of “no-

    cost” ensemble Kalman Smoother (EnKS) proposed by Kalnay et al. (2007) and tested by Yang et al. (2008). During the process, the observations from the past were used more than once by running ensemble Kalman Smoother.

Mozheng Wei:

    Showed slides and updated the EMC progress in estimating analysis error variances from GSI. The pre-conditioned conjugate gradient methods in GSI is used to deliver two sets of gradients vectors, one is gradient with respect to x (g)

    and the other is with respect to the transformed vector z (h) (subroutine

    pcgsoi.f90 was modified). These two sets form a bi-orthogonal system. The method consists of using sub-sets of these vectors to estimate the Hessian matrix. It capitalizes on relationship between conjugate gradient and the Lanczos algorithms for solving large scale eigenvalue problem. This method was first proposed by Fisher and Courtier (1995). The experiment has shown that with a small number of conventional obs data, the reductions of error variances in observation regions are identified with 30 dominant eigenvectors. It is still a challenge to convert the variances of stream function and velocity potential into the variances of u and v which will be more useful in ensemble. Craig Bishop’s

    suggestion of using 30 eigenvectors will be investigated. Another difficulty is to get accurate case-dependent background variances in order to get the accurate analysis error variances. One possible solution is to use the 6-hour ensemble

     4

    forecasts to construct background error variances. Another different direction of this work was also mentioned, that is to use the analysis data from different NWP centers.

    Craig Bishop:

    Sent some slides and two submitted manuscripts about his new adaptive error covariance localization for 4-dimensional ensemble data assimilation. This technique is called ECO-RAP (Ensemble COrrelations RAised to a Power). ECO-RAP is a new flow-adaptive localization method for ensemble DA. It raises ensemble correlations to a power to selectively reduce spurious correlations, and it adapts to changes in the propagation and scale characteristics of errors. The results show that the new method outperforms non-adaptive localization method when true errors are propagating or the error correlation length scale is varying.

Zoltan Toth:

    Summarized the activities and progress so far and planned to have the next meeting in June or July 2008. Groups discussed the future of Ens-DA activities. Craig Bishop suggested looking at observation data which need to be improved, especially satellite data bias correction needs to be improved. Mozheng Wei suggested to use a new GSI benchmark based on the new implementation of FOTO in May 2008. This view was shared by Jeff Whitaker and Tom Hamill.

     5

    SUMMARY OF TELECONFERENCE DISCUSSION

    October 4, 2007, 1:00pm 3:00 pm EST

    Present: J. Whitaker, T. Hamill and X. Wang (CDC), C. Bishop and Dan Hodyss (NRL), M. Zupanski (CIRA/CSU), Jeff Anderson (NCAR) I. Szunyogh and E. Kalnay(UMD), Y. Song, Z. Toth and M. Wei (NCEP/EMC) , E. Kostelich (Arizona State Uni.). Fuqing Zhang (Texas A&M Uni.). E. Kalnay (UMD) and Joe Tribbia (NCAR) unable to attend.

Agenda:

    1). Review of progress by each group (5-10 minute each group with slides

     if possible)

    2). Plans for the next 3-6 months by each group

    3). Longer term issues-expected benefits & computational costs of ens-DA

     and hybrid variational approach.

Craig Bishop:

    Sent many slides and showed the results from his new adaptive error covariance localization for 4-dimensional ensemble data assimilation. This technique is called ECO-RAP (Ensemble COrrelations RAised to a Power). ECO-RAP is a new flow-adaptive localization method for ensemble DA. It raises ensemble correlations to a power to selectively reduce spurious correlations, and it adapts to changes in the propagation and scale characterisitics of errors.

     nd part of the slides showed how to generate very large ensemble from smaller 2

    scale turbulence ensembles. The method together with ECO=RAP was tested in T119L30 model using LETKF, and produced promising results. In the near future, he can help Jeff Whitaker and I. Szunyogh to implement these techniques in their ensemble DA system..

Milija Zupanski:

    Sent slides showing how to compute the minimization in MLEF with non-differentiable observation operator based on 1-d Burgers model simulating a shock-wave. Further slides showed some preliminary results from MELF with

     6

    WRF and real observations. When it was applied to Hurricane Katrina, it showed the impact of observations comparing with no-DA experiments.

    He will continue to work on some computer issues and research on this minimization method.

Jeff Whitaker:

    New slides showed the results for May and June 2007 experiments and compared with GSI in the same model resolution (T126L64) using 64 ensembles. With more observations added, the gap between the GSI and LETKF is narrower than before. When AMSUA data are used, LETKF is basically the same as GSI in N. Hem. in terms of O-F statistics. In the Tropics, LETKF is slightly better for channels 14 and 15, but a little worse for channels 1, 2 and 3. In the S. Hem., GSI is a little better for channels 12, 13, 14 and 15.

    When Radiosondes are used, GSI is better for temperature O-F in most cases. In single observation experiments, it is shown that the localization scheme in LETKF needs to be improved further in comparison with the GSI.

    In terms of computational cost, GSI (300secs) is much more efficient than LETKF (3000secs) which runs faster than EnSRF (7000secs). The next possible step is to work with Xuguang Wang testing hybrid method, and with Criag Bishop and Milijia Zupanski to test their methods.

Istvan Szunyogh:

    New slides were presented and showed some results from assimilation of satellite radiances using cutoff-based data selection strategy. Bias correction is done by augmenting the state with bias parameters in the LETKF. The method was shown to improve the overall performance compared with the other methods. Digital filter has been used. He has plan to investigate the balance issues further in the future studies.

Fuqing Zhang:

    Showed his previous slides that demonstrated the advantages of EnKF over 3D-WRF. One new slide in this conference showed the comparison between GSI and EnKF in predicting hurricane Humberto (2007). The forecasts from EnKF are clearly better than those from GSI analysis in predicting the pressure and wind speed near-shore during hurricane Humberto. However, one should be careful when evaluating the results. Running a mesoscale forecast initialized with the coarser resolution global GSI may not offer a fair comparison of the capabilities of the 2 DA schemes compared.

     7

Xuguang Wang:

    Presented a hybrid ETKF-3DVAR data assimilation scheme for WRF. The hybrid system has shown larger improvement over data sparse regions, such as ocean. The flow-dependent ensemble covariance has shown the largest impact on the regions with sparse observations. The forecasts from the hybrid system with real observations also outperformed the forecasts from WRF 3DVAR system which is still under development (not as mature as GSI).

Eugenia Kalnay:

    Sent 2 sets of slides about sparse computations within LETKF, and studies about observation impact within the LETKF without adjoint. But she was unable to join. Material will be presented at the next tele-conference.

Jeff Anderson:

    Pointed out that a global version of WRF is available for use. He suggested to use different localization schemes for different places, different types of observations and different model variables.

Mozheng Wei:

    Updated the DA activity using ensemble at EMC. DA group at EMC is planning to use ensemble information to improve the background covariance matrix in GSI.

Zoltan Toth:

    Summarized the activities and progress so far and planned to have the next meeting in the first half of January 2008. Discussed future Ens-DA activities.

     8

    SUMMARY OF TELECONFERENCE DISCUSSION

    May 17, 2007, 1:00pm 3:00 pm EST

    Present: J. Whitaker and T. Hamill (CDC), C. Bishop and Dan Hodyss (NRL), M. Zupanski (CIRA/CSU), Jeff Anderson and Joe Tribbia (NCAR) I. Szunyogh and E. Kalnay(UMD), Y. Song, Z. Toth and M. Wei (NCEP/EMC) , E. Kostelich (Arizona State Uni.)

    Agenda:

    1). Review of progress by each group (5-10 minute each group with slides

     if possible)

    2). Plans for the next 3-6 months by each group

    3). Status of benchmark, discussion on comparison of results

    4). Longer term issues-expected benefits & computational costs of ens-DA

     and hybrid variational approach.

Jeff Whitaker:

    Reported the progress he has made. Modified the DA code for new GFS model, and new GSI will be used for forward operator (GSI has been implemented since May 1, 2007) in ensemble DA. New codes have been ported and debugging has started at T126. Satellite data have not shown much benefit for the ensemble DA analysis. They are looking at flow-dependent error covariance and model errors. Increasing model resolution is expected to reduce the role of model errors.

Eugenia Kalnay:

    Sent some slides and presentations from recent meetings. The experiments using Quasi-Geostrophic model showed that 40-member LETKF is better than 12-hour 4D-Var which is better than hybrid (3D/Var + 20 BVs), while 3D-Var and 24-hour 4D-Var are the worst and best respectively. Singular vectors and bred vectors are compared with the errors of the day. It was demonstrated that the errors of the day are most important for providing background error covariance in LETKF and 3D-Var.

Jeff Anderson:

     9

    Reported the newly released MPI version of DART using WRF model. Ran test with re-analysis observations. New inflation strategy, which is called adaptive inflation, greatly improved the analysis results.

Milija Zupanski:

    Still working with setting up computers. Was working on development of new ensemble localization scheme that will take time into account.

Istvan Szunyogh:

    Finished a final version of paper for Physica D that summarizes the results based on LETKF without satellite radiances. Implemented LETKF on NCEP model and produced some good results from using satellite data. Gave support to Jeff Whitaker’s experiments.

Craig Bishop:

    Sent some new slides to show the results from his new localization scheme where a SENCORP moderation function is constructed from equivalent potential temperature correlation filed smoothing in both horizontal and vertical using a Gaussian weighting function in Fourier phase space was performed. He showed the impact of using these moderation function in DA on forecast and analysis with NOGAPS ensembles. It is going to be tested in NVADAS system at NRL first. In the future it can be used in Whitaker’s ens-DA system.

Mozheng Wei:

    Introduced the DA activity at EMC. DA group at EMC is planning to use ensemble information to improve the background covariance matrix in GSI. The initial plan is to use the hybrid method which consists of both static 3D-Var background and flow-dependent ensemble based error covariance matrix. Ensemble group is hoping to get the estimate of analysis error variance from GSI for ensemble use in initialization.

Zoltan Toth:

    Planned the next meeting in August 2007. Future Ens-DA should work with regional, ocean and global ensembles. Two-way comparison with future 4D-Var

     10

Report this document

For any questions or suggestions please email
cust-service@docsford.com