Progress review

Presentations

Hans gave a 4D-Var presentation at the AFWA mid-year review. He thanks everyone for the input.

Hans gave a WRF 4D-Var presentation at NRL, Monterey.

Bug fix (Xin)

Xin removed a bug in JcDFI and it can now run with CG. Comparison between the JcDFI in QN and JcDFI in CG will be carried out. Once the correctness of JcDFI with CG, the QN will be removed again.

Xin also made the adjoint testing code in wrfvar, which tests the correctness of WRF_TL and WRF_AD. However the test has only 5 digits correct. The debugging effort will continue.

Physics (Qingnong, Jimy, Xiaoyan)

The nonlinear DU code has been tested and gives good results comparing to other sophisticated convection schemes.

Qingnong has collaborated with CAMS, and made the tangent linear and adjoint code with the CAMS colleagues. The automatically generated code uses memory, in contract to the TAF generated ones. Need more time to get used to the new TL/AD code.

The tangent linear code has passed the gradient test.

The adjoint test is on-going.

Meteorological testing (Yongsheng and Hans)

Reruns of the Haitang case are under evaluation. The one-sided 4dvar seems to give the best scores.

Yongsheng made experiments on two new cases: Shanshan and IHOP. Evaluation is on going.

Multi-incremental (Xin and Hans)

The multi-incremental version finally works with one outer-loop. Work has begun on the multi-outer-loops.

Dale asked whether the multi-incremental code can run in 3D mode. Hans agreed to test this.

John reminded the group to be careful with the wrfvar script. Hans said the multi-incremental related scripts are now all in the 4dvar branch. John also hope to include a 3/4D multi-incremental case in the regression test with some document.

Optimization (Tom)

Finished prototype code to reduce static load imbalance in the adjoint. Static load imbalance is reduced by shifting specified numbers of rows and/or columns between tasks on the domain edges and interior. This technique has proven useful in similar cases with other forecast models. Initial implementation is hard-coded in RSL_LITE for now and will be generalized and exposed as namelist options if tests show performance improvement. Testing is underway.

Other projects

Yongrun just started working on 4D-Var.

Qingnong agreed with U. Arizona to support a proposal using WRF adjoint model.

The group had a short discussion on the friendly-user support.

AOB

The next meeting will be 10:00 August 31, room 2072.