the setup in fv3jedi for pressure analysis (delp and surface pressure)

[TingLei-NOAA]

This issue is to facilitate further discussion/investigation of this issue as the title.

@delippi has found #507 (comment) that the delp in the analysis is not updated in his rrfs runs. It is further confirmed this occurs for some standard rrfs fv3-jedi setup as in the rdasapp ctest .
Further tracing this in fv3-jedi, my current understanding is that : not like in gsi, in which, after the analysis , the surface pressure increment would be used to update delp increment (by the hydrostatic relationship), fv3jedi does a more general ps to delp (and vice versa) in the variables changes in the control2analysis block. That means we have to use both delp and ps as the analysis variables and explicitly specify ps (no delp) as the active variables in the background error session.
The above setup (using both ps and delp as the analysis variables but only ps as the control variable) will not work for enkf . If we still want to only use ps (not delp aggressively directly analyzed), we have to add an gsi-similar like ps to delp conversion in fv3-jedi.

We had some discussion before as #456, in which I thought fv3-jedi was doing similarly on this as GSI does ,misleadingly.
Also bring this to @XuLu-NOAA 'a attention.
An clarification : "assignment" of the issues by me means the attention is hoped for and doesn't mean any responsibilities. :).
Update on Jan.12,2026: the "Fv3-JEDI-compatible " treatment of delp and ps in the fv3jedi hybrid analysis is done by : (1) both of ps and delp could (need) be used as analysis variables; in the static B part, only ps is used as the control variable (go through the B ) and the above mentioned linear relation is used between ps and delp; (2) for ensemble error part, ps and delp are analyzed by the localized ensemble sampled covariance. Hence, finally, this "F3-jedi compatible" analysis would give such a self-consistent hybrid analysis, in which, the ensemble component of ps increment would be just be discarded after the analysis. This is a major difference from what the current fv3reg GSI hybrid run does

[ShunLiu-NOAA]

Following our discussion today, the first step is to compare the Psfc-derived delp increments in GSI against the delp increments in JEDI. We can likely implement a Psfc assimilation strategy in JEDI that mirrors the GSI approach.

[TingLei-NOAA]

For discussion convenience, some relevant subroutines are listed here:
For jedi, ps_to_delp_tl/ad (defined in ./VariableChange/Utils/pressure_variables_mod.f90 and used (if applicable) in fv3jedi_varcha_c2a_mod.f90 .
For gsi used by rrfs (fv3 model), to map the ps increment to delp is done in wrfv3_netcdf in gsi_rfv3io_mod.f90 as

      allocate(g_prsi(lat2,lon2,nsig+1))
      allocate(ges_ps_inc(lat2,lon2))
      ges_ps_inc=ges_ps-ges_ps_bg
      g_prsi(:,:,nsig+1)=eta1_ll(nsig+1)
      do i=nsig,1,-1
        g_prsi(:,:,i)=ges_delp_bg(:,:,i)+g_prsi(:,:,i+1)  
      enddo
      do i=1,nsig+1
        g_prsi(:,:,i)=g_prsi(:,:,i)+eta2_ll(i)*ges_ps_inc
      enddo
      do i=1,nsig
        ges_delp(:,:,i)=g_prsi(:,:,i)-g_prsi(:,:,i+1)
      enddo
      ges_delp=ges_delp*1000.0_r_kind

[delippi]

After using @TingLei-NOAA 's yaml modifications to enable delp increments in jedi (left), I compare them against the increments of delp from gsi (right). It seems the current method in jedi does match gsi fairly well as it is.

Edit: Just noticed in my screen capture, I made a typo in setting the colorbar bounds the same which is a bit misleading. jedi's colorbar was mistakenly set to [-1.2, 1.0] while gsi was using [-1.2, 1.2]

[TingLei-NOAA]

@delippi That is great! That behavior is what we expected.
Really appreciate your insights that helps a lot to sort out this issue.

[HuiLiu-NOAA]

I see the patterns seem similar, but there seems a systematic bias between the two, that is, the top plot has a positive bias vs. the bottom one. Another plot showing the actual diff. would be helpful.