TRACER-MIP: regional nested coastal Tier 1 control

This page records the verified protocol of the TRACER model intercomparison (Houston, 17 June and 7 August 2022), the NumericalEarth capabilities BreezeLab builds on, what is implemented, and every material departure from the reference treatments. It is a working record, not a claim of a completed MIP submission.

Pinned sources (cases/tracer_mip/inputs.toml, checksum-verified): Roadmap.md, the March-2025 roadmap PDF and the aerosol-profile notebook of ARM-Synergy/tracer-mip @ 416423ac. The numbers below were re-read from those files; cases/tracer_mip/protocol.toml holds the transcription that the code reads (tracer_mip_protocol() validates it).

Protocol (Roadmap Tables 1-3)

Item	Protocol value
Cases	24 h from 06 UTC 2022-06-17 and 06 UTC 2022-08-07
Initial and boundary data	3-hourly ERA5 0.25° pressure-level reanalysis
Nests	2, one-way only, common center 29.4719 N, 95.0792 W; polar stereographic "or similar"
Grid-1	750 × 750 at 2000 m, Δt = 3 s, full output every 60 min
Grid-2	500 × 500 at 500 m (the innermost ACPC-MIP nest), Δt = 1.5 s, full output every 10 min and every 2 min from 17 UTC to 01 UTC for cell tracking
Vertical	95 ACPC scalar levels (m AGL, first level -24 m), top ≈ 22 km / 50 hPa
Physics	Coriolis on, no convection scheme, interactive land surface (urban physics if available), two-moment or bin microphysics with prognostic droplet number, best PBL/diffusion every step, LW+SW radiation every 60 s, radiatively inactive aerosols
Aerosol (Tier 1)	fixed at each step; two lognormal modes: Aug 7: 1425 cm⁻³ (Dm 49 nm, σ 1.8) + 263 cm⁻³ (175 nm, 1.4); Jun 17: 3443 cm⁻³ (30 nm, 1.5) + 531 cm⁻³ (136 nm, 1.5); converted with ρ = 1.159 kg m⁻³ to mg⁻¹; lidar-derived shape, total floor 50 mg⁻¹, floor only above 6 km; κ = 0.26
Variants	×3 and "3× lower (coefficient 0.3)" surface numbers, same shape; the roadmap also writes ⅓ — unresolved for LOW
Required outputs	3-D state and water species with numbers, 2-D surface/radiation fields, and separate process rates (condensation, evaporation, deposition, sublimation, melting, freezing, droplet and ice nucleation, riming of cloud and rain, autoconversion + accretion) as kg/kg per output interval

NumericalEarth capability inventory (verified 2026-10-07)

The ERA5, Atmospheres, NestedModels and Lands modules and the examples breeze_downscaling_era5.jl, era5_forced_slab_land.jl and breeze_over_slab_land.jl live in NumericalEarth.jl, not in Breeze. Pinned revision d07eb240a5679b7a6913e4d979e8378e76786ad7 (v0.8.1, 2026-10-06): its compat (Breeze = "0.11", Oceananigans = "0.111-0.113") accepts the project pins Breeze 5264d3c (0.11.3) and Oceananigans 847e125 (0.113.3).

Available: an ERA5 pressure-level PrescribedAtmosphere parent on the native 0.25° grid with time-varying geopotential heights; nested_atmosphere_model(child_grid, ERA5HourlyPressureLevels(); dates, dir, terrain = ETOPO2022(), …) that builds a Breeze compressible split-explicit child on a terrain-following LatitudeLongitudeGrid with open lateral boundary conditions, Davies relaxation, a ρw lid/wall sponge, ERA5 initialization and adiabatic balancing; SlabLand (prognostic skin temperature, bucket or variably-saturated hydrology) coupled through AtmosphereLandModel with Monin-Obukhov fluxes and RRTMGP surface radiation; PrescribedOcean for SST-driven atmosphere-ocean fluxes.

Missing or mismatched for this protocol:

Nesting a Breeze child in a Breeze parent is not implemented (the parent must be a PrescribedAtmosphere); one-way 2 km → 500 m nesting is therefore done offline: the outer run writes its prognostic state over the inner region and a PrescribedAtmosphere is built from that output.
Land and ocean fluxes under one Breeze atmosphere are "at most one surface type per cell"; the coastal domain needs a mask (BreezeLab pins sea cells of the slab to ERA5 SST with a saturated bucket and sea roughness).
The extension imports Breeze.AtmosphereModels.surface_precipitation_flux, renamed bottom_precipitation_flux before the pin; Julia warns and falls back to a zero rain flux into the land bucket.
P3 AerosolActivation modes are vertically uniform; Tier-1 height-dependent prescribed aerosol needs a scaled activation.
No accumulated process-rate diagnostics; BreezeLab adds them.
NumericalEarth's weak-dependency compat forces RRTMGP 1.0.1 → 0.22.3 when added to the BreezeLab environment, so it is pinned only in the case sub-environment cases/tracer_mip/Project.toml.

BreezeLab implementation choices and departures

Horizontal grid: Oceananigans LatitudeLongitudeGrid instead of polar stereographic. Extents are chosen so that cells are 2 km × 2 km (outer) and 500 m × 500 m (inner) at the center latitude: outer 22.727-36.217 N, 102.827-87.332 W; inner 28.348-30.596 N, 96.370-93.788 W (inside the roadmap's corner envelope). Zonal spacing varies from 2.12 km (south) to 1.85 km (north) across the outer domain.
Vertical grid: the 95 scalar levels are RAMS zt heights; interfaces are their midpoints, the bottom face is 0 m and the top face extrapolates to 22181 m: 94 cells, Δz = 50 m at the surface, 300 m above ≈ 5.5 km (acpc_vertical_faces).
Aerosol: exact port of the notebook (tracer_mip_aerosol_profile), tested against its printed tables. The bulk κ = 0.26 is imposed by setting the P3 solute parameter β_act = κ (van 't Hoff factor derived from the chosen density/molar mass).
Microphysics: Breeze P3 (two-moment ice with predicted properties, prognostic droplet number). Not a bin scheme.
Land: SlabLand (skin temperature + bucket) is a slab, not a multi-layer soil/vegetation/urban model; initialized from ERA5(-Land) skin temperature and soil moisture. No urban physics.
Radiation: RRTMGP all-sky every 60 s; aerosols are radiatively inactive (protocol).
Boundaries: hourly ERA5 (the protocol minimum is 3-hourly), interpolated linearly in time; Davies relaxation over the outermost cells; no interior large-scale tendency is applied (forcing enters only through the boundaries and the surface).
Turbulence/PBL: TKEBasedTurbulenceClosure (vertical eddy diffusivity with prognostic TKE, vertically implicit) is the control's closure. The first 24-h control (job 233) ran without a closure and shows a surface-layer defect (runs/outer_era5_control_aug07_job233/diagnostics/surface_wind_check.{toml,png}): the lowest-level (25 m) wind, initialised at ERA5's 1000 hPa value (domain mean 7.2 m/s vs ERA5 10 m 4.9 m/s — the ERA5 pressure levels carry no surface layer), falls within one hour to 2.6–3.0 m/s, 0.58–0.83 of ERA5's 10 m wind for the rest of the run, while the second level (76 m) keeps the free-stream value (site profiles: 3.1 vs 8.0 m/s at 3 h, 1.9 vs 4.4 m/s at 9 h, with ERA5 5.3 and 3.8 m/s at ≈160 m). The surface stress is confined to the 50 m lowest layer because nothing mixes momentum vertically; the Monin-Obukhov fluxes then see an unphysically weak wind. Judged a defect of the no-closure configuration; the control is rerun with the TKE closure. SmagorinskyLilly cannot be used on 50 m × 2 km cells: its isotropic filter width (≈585 m) makes the explicit vertical diffusion unstable (νΔt/Δz² > 1/2) in the first step. A horizontal-only Smagorinsky with the 1-D TKE closure would be the LES-consistent mesoscale choice, but the pinned Breeze cannot combine closures (post-pin #1036); horizontal mixing is the WENO scheme's.
Rain to land: NumericalEarth imports the renamed surface_precipitation_flux; the extension defines it at run time in Breeze.AtmosphereModels as the sign-flipped bottom_precipitation_flux so the child's rain reaches the bucket.
Nesting: one-way and offline — the outer run saves its state on the inner region + halo every 10 min, and outer_run_parent rebuilds a PrescribedAtmosphere on a static 3-D-height grid for tracer_mip_inner_simulation.

Runs so far (all on wpcluster; see the status file for job IDs)

Run	Configuration	Result
CPU smokes (jobs 204–215)	real ERA5, 32×32×94, 36 s	stable without closure and with TKE closure; Smagorinsky fails in step 1
GPU pilot (job 217)	real ERA5, 750×750×94, 1 h, A100-80GB	completed; 40.3 GB; 4.3 s wall per 3 s step (≈86 min per simulated hour)
GPU control (job 233)	same, 24 h, no closure	running; surface-layer wind defect diagnosed (see Turbulence/PBL)
GPU control (job 339)	same, 24 h, TKE closure	queued after job 233 (replacement August 7 control)
Inner-nest CPU test (job 227)	parent from the pilot's saved state, 8×8×24 child, 6 s	passes

Status

See plans/status/TRACER-MIP.md in the coordination checkout for the live status, job IDs and results. Implemented and tested on CPU: protocol metadata, grids, vertical faces, aerosol profiles and P3 modes, the Tier-1 prescribed aerosol in P3, process-rate accumulators, ERA5 staging, the outer-domain constructor with coastal exchange, and the offline inner nest. The 24-h August 7 outer control is running; the inner 500 m run, its 10-min/2-min writers and the June 17 case are next. A pilot is not a completed MIP; nothing here has been compared against observations.