Physical calibration and ablations

apps/mars-calibration differentiates coupled GCM trajectories with three bounded physical controls:

Control	Forcing field	Meaning
CO₂ longwave	ames_co2_longwave_opacity_scale	Gas thermal optical-depth multiplier
Dust longwave	ames_dust_longwave_opacity_scale	Prescribed dust thermal optical-depth multiplier
Surface exchange	surface_exchange_multiplier	Consistent heat/momentum bulk exchange multiplier

These controls do not change composition or gas constants. Short-window loss reduction is not evidence of equilibrated calibrated climate.

rtk proxy uv pip install -e ./apps/mars-calibration
rtk proxy .venv/bin/python scripts/run_mars_paper_experiments.py --help
rtk proxy .venv/bin/python scripts/run_differentiable_experiments.py --help

The frozen input bundle contains restart, targets, forcing and checksums. outputs/nautilus-calibration-inputs in existing examples is a prepared local artifact, not a dataset shipped with a fresh checkout. Prepare it using the application README in the repository.

Experiment protocols

The original driver compares L-BFGS-B/Powell with fixed oracle budgets and separate finite-difference checks. The paper runner adds paired full-physics and operator-removal cases. Its annual configuration spans 59,356,800 seconds, 197,856 steps at 300 s; snapshots are instantaneous and RMSE is a sample average.

The newer differentiable runner offers gradients, calibration, neural, prepare-synthetic, and ablations. Physical fitting uses forward-mode AD; neural fitting uses reverse-mode AD and bounded heating. It uses Adam and norm clipping, so its update budgets differ from the original oracle budgets.

The experiment runbook provides input preparation, alignment, gradient checks and reporting commands. Use fresh output directories. Inspect acceptance fields and source hashes; completion alone does not establish scientific success. --require-gpu rejects CPU fallback.

Checkpointed climate integration

With the named TES and Ames inputs staged, start with a one-sol pilot in a fresh directory before running this substantial workload:

rtk proxy .venv/bin/python scripts/run_gcm3d_ablation.py \
  data/tes/mgs_tes_surface_1deg.nc \
  --output-dir outputs/climate-baseline --config convection \
  --truncation T21 --layers 12 --dt 300 --diurnal --initial-ls 0 \
  --sols 2006 --chunk-sols 5 --cooldown-seconds 0 \
  --co2-lw-scale 0.25 --dust-lw-scale 0.25 \
  --surface-exchange-multiplier 1.0 \
  --ames-dust-reference data/ames/fv3betaout1/ames_surface_reference.nc

convection is the cumulative full-physics configuration. Resume a matching run with --resume. Explicit 0.25 scales match the corrected baseline; script defaults are 1.0. Check phase-matched repeatability, including deep soil, rather than simply counting elapsed years. See validation.