Improve Initialization from ERA5 and MSLP Diagnostic #4098

costachris · 2025-11-12T23:41:03Z

Improvements to WeatherQuest ERA5 initial conditions:

Adjust for ERA5/ClimaAtmos topo differences when initializing from surface pressure.
Improve initialization by including upstream 3D lat/lon/z interp, extrapolation beyond 1000mb surface, and high resolution custom vertical grid. These changes are needed to reduce vertical interpolation errors because upstream Clima utilities lack tri-linear interpolation to the spectral element space, resulting in what were otherwise significant interpolation errors.
Account for lapse rate in mslp sea level reduction (diagnostics).

The colorbars vary, but note each improvement significantly reduces the RMSE/Bias.
Current 😭

Topography correction in surface pressure init 👀

Topography correction + account for lapse rate in MSLP diagnostic 🥳

Julians42 · 2025-11-14T04:26:35Z

src/utils/weather_model.jl

    @assert all(map(x -> x in (keys(ncin)), in_dims)) "Source file $source_file is missing subset of the required dimensions: $in_dims"

    # assert ncin has required variables
    req_vars = ["u", "v", "w", "t", "q", "skt", "sp"]


If it breaks without surface_geopotential catch it with the @assert here

src/utils/weather_model.jl

Julians42 · 2025-11-14T04:38:34Z

src/utils/weather_model.jl

+        end
+        var_obj = defVar(ncout, dst_name, FT, ("lon", "lat", "z"), attrib = var_attrib)
+        # Read first time slice and coalesce; follow same convention as sp (use [:, :, 1])
+        data2d = FT.(coalesce.(ncin[src_name][:, :, 1], NaN))


Do we need a coalesce? Are there instances when the data is NaN and our model will still run? If not then we could consider asserting not NaN

src/initial_conditions/initial_conditions.jl

src/diagnostics/core_diagnostics.jl

Julians42 · 2025-11-14T05:10:28Z

src/diagnostics/core_diagnostics.jl

+    bracket = similar(t_level)
+    @. bracket = max(ϵ, 1 + Γ * z_level / t_level)
+
+    # exponent = g / (R_m Γ)
+    exponent = similar(R_m_surf)
+    @. exponent = g / (R_m_surf * Γ)


This will allocate new arrays which is slow. Maybe @dennisYatunin could recommend a fix?

I consolidated the logic and reduced the number of new variables. But also diagnostics are computed on the CPU so I'm not sure we need to worry about allocations as much here (note we already materialize a lot of our lazy fields throughout diagnostics as is). But maybe @dennisYatunin can comment how much it matters in the end.

src/diagnostics/core_diagnostics.jl

Julians42

Left some comments to improve clarity and then one performance optimization we should do in the diagnostics for Dennis. Glad this is looking so much better!

…re. Account for lapse rate in mslp sea level reduction (diagnostics).

costachris requested a review from Julians42 November 12, 2025 23:48

costachris force-pushed the cc/improve_wxmodel_pressure_init_diag branch from 6f20da6 to 1ce4418 Compare November 13, 2025 00:47