Met Driver Intercomparison Tutorial

One of the key design features of AtmosTransport.jl is that the same model code runs with different meteorological driving data simply by pointing to a different TOML configuration file. This enables systematic studies of how transport uncertainties from meteorological reanalyses propagate into tracer simulations and flux inversions.

Motivation

Quantifying met driver uncertainty is essential for inverse modeling of greenhouse gas fluxes. Differences in winds, boundary-layer mixing, and convection between ERA5, MERRA-2, and GEOS-FP propagate directly into:

• Simulated tracer distributions (XCO₂, XCH₄)
• Retrieved surface fluxes via 4D-Var
• Inter-hemispheric exchange times
• Vertical profiles at validation sites

With legacy models (TM5 uses only ECMWF; GEOS-Chem uses only GEOS), disentangling model errors from met driver errors requires running two entirely different codebases. AtmosTransport.jl eliminates this confounding factor.

Example: Three-Way Met Driver Comparison

The workflow below runs the same tracer transport simulation with three different meteorological data sources and compares the results.

Step 1: Define common model setup

using AtmosTransport
using AtmosTransport.Grids
using AtmosTransport.IO: default_met_config, build_vertical_coordinate
using Dates

# Build vertical coordinate from the GEOS-FP config (72 levels);
# ERA5 would give 137 levels — choose whichever matches your comparison.
config = default_met_config("geosfp")
vert = build_vertical_coordinate(config; FT=Float64)

# Common grid — all met drivers will be regridded to this
grid = LatitudeLongitudeGrid(CPU();
    FT   = Float64,
    size = (360, 180, n_levels(vert)),
    longitude = (-180, 180),
    latitude  = (-90, 90),
    vertical  = vert)

# Common tracer setup — uniform initial condition (NamedTuple of 3D arrays)
Nx, Ny, Nz = 360, 180, n_levels(vert)
c = fill(Float64(400.0e-6), Nx, Ny, Nz)
tracers = (; CO2 = c)

# Common time window
t_start = DateTime(2024, 3, 1, 0, 0, 0)
t_end   = DateTime(2024, 3, 8, 0, 0, 0)   # one week

Step 2: Run with each met driver

The workflow for each met source follows the same pattern used in scripts/run_era5_edgar_forward.jl:

1. Load met data for each window (3-hourly or 6-hourly)
2. Precompute mass fluxes via compute_air_mass! and compute_mass_fluxes!
3. Run strang_split_massflux! for TM5-faithful advection

using AtmosTransport.Advection: build_geometry_cache, allocate_massflux_workspace,
    compute_air_mass!, compute_mass_fluxes!, strang_split_massflux!

results = Dict{String, Any}()

for met_name in ["geosfp", "merra2", "era5"]
    cfg = default_met_config(met_name)
    vc  = build_vertical_coordinate(cfg; FT=Float64)

    g = LatitudeLongitudeGrid(CPU();
        FT=Float64, size=(360, 180, n_levels(vc)),
        longitude=(-180, 180), latitude=(-90, 90), vertical=vc)

    # Allocate arrays (zero-allocation inner loop)
    Δp  = zeros(Float64, 360, 180, n_levels(vc))
    m   = similar(Δp)
    am  = zeros(Float64, 361, 180, n_levels(vc))
    bm  = zeros(Float64, 360, 181, n_levels(vc))
    cm  = zeros(Float64, 360, 180, n_levels(vc)+1)
    gc  = build_geometry_cache(g, Δp)
    ws  = allocate_massflux_workspace(m, am, bm, cm)

    run_tracers = (; CO2 = copy(c))  # deep copy per met source

    # ... per-timestep loop: load met → compute fluxes → strang_split_massflux! ...

    results[met_name] = copy(run_tracers.CO2)
end

Step 3: Compare results

Key diagnostics to examine:

using Statistics: mean

# Global mean mixing ratio
for (name, field) in results
    println("$name: global mean = $(mean(field)) ppm")
end

# Zonal mean cross-sections
for (name, field) in results
    zonal = mean(field, dims=1)  # average over longitude
    # Plot latitude-altitude cross-section...
end

# Differences
diff_geos_merra = results["geosfp"] .- results["merra2"]
diff_geos_era5  = results["geosfp"] .- results["era5"]
diff_merra_era5 = results["merra2"] .- results["era5"]

What to Compare

Horizontal Transport (Winds)

• Zonal mean u, v differences
• Implications for inter-hemispheric exchange
• Storm-track representation

Vertical Transport (Omega)

• Tropical upwelling (Brewer-Dobson circulation)
• Subsidence patterns
• Critical for CO₂ seasonal cycle amplitude

Boundary Layer Mixing (Kz / kh)

• PBL height differences
• Mixing efficiency near the surface
• Directly impacts surface flux inversions — this is often the dominant source of transport error for ground-based measurements

Deep Convection (Mass Fluxes)

• Tropical convective redistribution
• Critical for CH₄ and CO vertical profiles
• Often the most poorly constrained transport process

Surface Pressure

• Mass budget consistency across reanalyses
• Affects column-average quantities (XCO₂)

Regridding Considerations

The three met sources have different native grids:

When comparing, either:

1. Regrid to a common grid (e.g., 2° × 2°) using conservative regridding
2. Compare area-weighted statistics (global means, zonal means, vertical profiles)

The model's Regridding module handles both conservative and bilinear regridding. See src/Regridding/Regridding.jl.

Extension: Adjoint Sensitivity to Met Drivers

The adjoint model can quantify how met driver differences affect the gradient of a cost function (e.g., in a CO₂ flux inversion):

# Each forward operator (advect_x_massflux!, etc.) has a paired adjoint;
# running the adjoint with different met drivers quantifies sensitivity.
for met_name in ["geosfp", "merra2", "era5"]
    # ... run adjoint_strang_split_massflux! with same met data ...
    # Compare adjoint gradient maps across met drivers...
end

This provides a rigorous framework for estimating "transport model uncertainty" in flux inversions — a quantity that is typically estimated ad hoc by comparing TM5 and GEOS-Chem results (confounding model structural differences with met driver differences).

This page was generated using Literate.jl.