sbc: rotate initial wind/stress coefficients on rotated grids

[koldunovn]

Summary

On a rotated grid (rotated_grid = .true., e.g. the standard CORE2 setup), the initial surface-forcing wind/stress interpolation coefficients are never rotated into the rotated-grid frame. As a result, for the first forcing window of a run the wind stress applied to both the ocean and the sea ice stays in the geographic frame and is mis-directed by the local geo→rotated (g2r) angle.

Mechanism

The g2r rotation of the wind/stress is applied only in sbc_do (src/gen_surface_forcing.F90), and only when the time-interpolation coefficients are refreshed:

! sbc_do: do_rotation_wind is set ONLY inside the getcoeffld (refresh) branch
if ( ((rdate > nc_time(t_indx_p1)) .and. ...) .or. force_newcoeff) then
   call getcoeffld(fld_idx, rdate, partit, mesh)
   if ((l_xwind .and. (fld_idx==i_xwind)) .and. rotated_grid) do_rotation_wind=.true.
endif
...
if (do_rotation_wind) then        ! rotate coef_a/coef_b for wind
   ... call vector_g2r(coef_a(i_xwind,i), coef_a(i_ywind,i), ...) ...
end if

The cold-start initialization nc_sbc_ini loads the very first set of coefficients but never rotates them:

! nc_sbc_ini
do fld_idx = 1, i_totfl
   call getcoeffld(fld_idx, rdate, partit, mesh)   ! load first coefficients
end do
call data_timeinterp(rdate, partit)                ! build atmdata -> NO rotation

At cold start, force_newcoeff is .false. (no year change) and rdate lies inside the first forcing interval, so the refresh condition in sbc_do is not met on the first steps → do_rotation_wind stays .false. → the rotation block is skipped → the wind used (atmdata = rdate*coef_a + coef_b) is the unrotated geographic wind. The rotation only kicks in once the run crosses the first forcing-window boundary and sbc_do calls getcoeffld again.

Impact

• Transient: it self-corrects after the first forcing-window crossing (for 3-hourly JRA55 at dt = 1800 s, ≈ 6 steps / ≈ 3 model hours).
• Magnitude-preserving: a rotation is orthogonal, so |stress| is unchanged. This is why the bug is invisible to every magnitude-based diagnostic — u*, turbulent fluxes, SST/SSS/SSH, ice area/volume, ice/ocean speed ratios — and to multi-year climate validation. Only the wind-stress direction is wrong, so it perturbs the initial ice drift and surface-current direction for the first few steps.
• Scope: only rotated grids (rotated_grid = .true.). Non-rotated setups are unaffected.

How it was found

Discovered while bit-comparing a C re-implementation of the FESOM2 forcing/EVP path against this Fortran code at ocean step 1 (CORE2 mesh, JRA55, dt=1800, cold start). Every EVP input was bit-identical except the wind-on-ice stress stress_atmice, which differed at all surface nodes. The difference was a pure rotation: per-node |stress| matched to 1.0000, while the direction differed by exactly the local g2r vector angle (stress_C = g2r(stress_Fortran), residual 0.0 at every node). Tracing it showed the C code rotated the wind from step 1 while this Fortran code did not, isolating the missing rotation to nc_sbc_ini.

Fix

In nc_sbc_ini, after loading the initial coefficients, apply the same g2r rotation that sbc_do already applies on refresh (guarded by l_xwind/l_xstre and rotated_grid). It is a faithful mirror of the existing sbc_do rotation block, so steady-state behaviour is unchanged; it only fixes the cold-start window.

Validation

Compared step-1 wind-on-ice stress (CORE2, JRA55, dt=1800, cold start) against a rotated-frame reference, before vs after the fix:

quantity	before fix	after fix
stress_atmice C-vs-ref, per node	pure rotation: `	·
resulting first-EVP-subcycle ice velocity	~70% relative diff	~4e-16

Multi-year climate is unchanged (the bug was a bounded transient that self-corrects after the first forcing window), confirming the fix only touches the initial-window forcing direction.