Hello to all,
I am trying myself to reproduce several models. This serves me to understand.
However, I am confused about the ADJOINT and TANGENT models; the files roms_iad.nc and roms_itl.nc are needed and I have been trying to find bibliography that explain what they are and how they are built.
I wonder if someone could point me into a direction where I can find these issues and read about them?
There are several questions I have, for example, if these files are built "one the fly" and re-read or if they belong to different runs or if they are built with different means.
All suggestions or examples, will be very appreciated.
Thanks!
Adjoint and tangent input files
- arango
- Site Admin
- Posts: 1368
- Joined: Wed Feb 26, 2003 4:41 pm
- Location: DMCS, Rutgers University
- Contact:
Re: Adjoint and tangent input files
The ITLNAME and IADNAME initialization NetCDF files are needed for specific algorithms like adjoint sensitivities, observation impacts and observation sensitivities. To generate those files you will need to know how to write the adjoint for the specific functional in ADSNAME. They are very advanced algorithms and require knowledge. We have written several publications describing these algorithms. Check my ResearchGate profile for the list of those publications.
In 4D-Var, the TLM and ADM kernels are initialized from zero, but the observations misfit forces them into the iterative inner loops. We have examples for the WC13 application in the test repository, and some of them are described in WikiROMS 4D-Var Tutorials. Moreover, the 4D-Var algorithm will generate ITLNAME internally but IADNAME is not needed because the ADM always starts from zero. The adjoint dynamics enter from the forcing operators.
There are a couple of ways how to specify the functional sensitivity by the delta functions in the ITL and IAD for the state variables of the specific metric or directly by generating the ADSNAME. There are examples in WC13. For example, check the Matlab script adsen_37N_transport.m.
In 4D-Var, the TLM and ADM kernels are initialized from zero, but the observations misfit forces them into the iterative inner loops. We have examples for the WC13 application in the test repository, and some of them are described in WikiROMS 4D-Var Tutorials. Moreover, the 4D-Var algorithm will generate ITLNAME internally but IADNAME is not needed because the ADM always starts from zero. The adjoint dynamics enter from the forcing operators.
There are a couple of ways how to specify the functional sensitivity by the delta functions in the ITL and IAD for the state variables of the specific metric or directly by generating the ADSNAME. There are examples in WC13. For example, check the Matlab script adsen_37N_transport.m.
Re: Adjoint and tangent input files
Thank you very much, that is a very clear answer. I will follow it.
all best.
all best.