Model, Ames, MCD and ARCO comparison

Generate an offline browser report from the staged reference files:

rtk proxy .venv/bin/python -m cli.main mars compare \
  --config cli/configs/reference-comparison.json \
  --output outputs/reference-comparison-new

The equivalent installed command is tform mars compare. The output directory must be new. Open its index.html in a browser; it needs no server or external JavaScript. The existing generated report is outputs/reference-comparison/index.html.

The report includes a season selector (Ls 45, 135, 225, 315 degrees), four-source surface-field grids, model-minus-reference grids, MOLA elevation, configuration and sampling tables, Gaussian-area-weighted metrics, CSV tables, input hashes, and AmesCAP NetCDF/template exports. Each field has the same color limits across sources; difference maps use shared symmetric limits centered at zero.

Meaning of the comparisons

The included configuration uses the existing physical baseline. It does not use or imply results from the separate 0.25-sol calibration. That calibration writes parameters and losses, not a seasonal climate dataset. To compare calibrated climate, freeze the selected parameters, run a separate physical evaluation with the same averaging protocol as the baseline, export its fields, and add it to a new report configuration. An improvement in the short calibration loss does not establish equilibrium or held-out climate improvement.

The sources have different scientific roles:

Source	Role	Current sampling
Physical GCM	Our deterministic model	Sparse instantaneous checkpoints averaged in a seasonal window
Ames FV3	Independent model reference	Duration-weighted five-sol averages
MCD 6.1	Model-derived climatology	Four cached fixed-local-time samples at a nearby season
ARCO-MACDA	Reanalysis	MY34–35 daily means in the seasonal window

These are diagnostic comparisons. Exact local-time/averaging alignment remains unresolved, winds compare different effective heights, dust scenarios differ, and MCD frost is a monthly statistic. MCD's four local times are a sampled mean, not an exact continuous daily mean. Original NetCDF attributes and input hashes remain in report.json. MOLA contours provide common geographic context; they do not assert identical terrain across the source models.

The parameter table distinguishes framework-specific calibration multipliers from comparable output quantities. Ames/MCD/ARCO multiplier entries remain unavailable rather than being guessed. Model multipliers in the supplied config are declared baseline values from the progress report, not recovered from each seasonal file.

Configuration contract

Paths resolve relative to root, itself relative to the configuration file. The first source is the model used in every difference and metric. Each source must declare its name, role, sampling, vertical level, dust scenario and paths. Patterns may include {season:03d} and glob wildcards. Duplicate source names are rejected.

For time-dependent inputs, declare season_variable (a one-dimensional variable on time), mean_dims: ["time"], and any explicit isel choices such as {"lev": -1}. Files are selected within the configured circular Ls half-width. All selected daily/time samples receive equal weight; unequal-duration averages must be preprocessed with explicit duration weights, as done by the existing Ames evaluator. No last-time or last-level choice is made implicitly.

The adapter supports the staged surface-temperature, pressure, wind-speed and frost schemas. It converts frost mass (kg/m²) to pressure equivalent with Mars gravity, accepts documented equivalent velocity-unit spellings, and rejects unknown or missing units. ARCO wind speed is formed from each daily vector before averaging those speeds. Missing variables are labeled unavailable; unresolved dimensions and non-finite comparison fields are errors.

Metrics are bias, MAE, RMSE, spatial correlation and model/reference means. Gaussian latitude quadrature is used on the T21 grid. Correlation is null/N/A for constant fields. Reference interpolation is periodic in longitude.

AmesCAP presentation

AmesCAP is NASA's Community Analysis Pipeline. Its native workflows target Ames GCM output. Our export adapter supplies already-selected two-dimensional latitude/longitude NetCDF fields for the model, Ames, MCD and ARCO, plus the common MOLA field zsurf. It does not invent time series, pressure levels, or native FV3 coordinates.

Each season contains:

amescap/ls045/
  comparison.in
  source1/comparison.nc   # physical GCM
  source2/comparison.nc   # Ames
  source3/comparison.nc   # MCD
  source4/comparison.nc   # ARCO

comparison.in uses MarsPlot's multi-simulation @N notation, HOLD ON panels, shared field ranges, difference expressions, and terrain contours. Paths in the template are absolute: regenerate or edit them after moving the report.

Install CAP in a separate environment according to its official instructions, then run from a directory where CAP may write its plots/ folder:

rtk proxy MarsPlot /absolute/path/to/report/amescap/ls045/comparison.in -ftype png --debug

Older CAP installations expose MarsPlot.py instead of MarsPlot and use -o png instead of -ftype png. Use PNG to avoid a dependency on system PDF/ghostscript tooling. The report's own PNGs and HTML do not require CAP at all. CAP is a presentation layer: numerical metrics are computed by the CLI before export. Cite AmesCAP separately when publishing its figures, along with Ames, MCD, ARCO-MACDA/MACDA and MOLA data sources.

The native Ames archive can also be processed with MarsInterp, MarsVars and MarsPlot; the CLI's existing scripts/benchmark_ames.py, scripts/benchmark_mcd.py, and scripts/evaluate_arco_seasonal.py remain the source-specific evaluation tools. Full pressure-level profile and zonal plots need matching vertical coordinates and separate exports; the new common grid covers the shared surface diagnostics only.

Verification in this workspace

The CLI generated all four seasonal reports and 48 field/reference metric rows. The generated Ls45 template rendered all 28 panels (eight PNG pages) using MarsPlot 3.3 from AmesCAP commit f3d23713566285709ade7f70198b3448c445b276, with temporary NumPy 1.26.4 and netCDF4 dependencies. Its plots are in outputs/reference-comparison/amescap-rendered/plots/. The project environment was not downgraded.

Current AmesCAP (MarsPlot 3.5, checked commit df04a413a9a228fe8aa31b3dccc3acb6c8f5b442) uses -ftype png and requires ~/.amescap_profile from the official installation instructions. That user configuration is absent here, so the current-release renderer was not fully validated. It can exit zero after reporting a missing profile; verify that actual plot files were produced. The successful render above is explicitly the 3.3 compatibility test, not a claim that all CAP releases were tested.