Skip to content

Frozen temperature-only postprocessing

This workflow fits a correction to cached physical forecasts. Unlike a neural tendency inside the solver, fitting does not differentiate through GCM integration or change wind, pressure, frost or soil.

The native ARCO-MACDA contract selects 112 training, 32 validation and 32 test starts with fixed seasonal coverage, observed initial fields, past-only soil initialization and exact native time alignment. Statistics and controls are fitted on training data only.

Workflow and acceptance

  1. Stage windows and create the frozen manifest/contract on CPU.
  2. Cache 144 training/validation forecasts on selected GPUs.
  3. Release GPUs, fit constant/ridge/neural controls on CPU, and freeze models.
  4. Cache 32 test forecasts only after passing the validation gate.
  5. Generate and verify reports, predictions, model hashes and budget records.

Follow the launch and acceptance runbook for complete commands, GPU requirements, failure and resume behavior. The cache launcher rejects CPU fallback; synthetic tests, staging and fitting can run on CPU.

Validation requires at least 10% RMSE reduction against physical forecasts and better RMSE than ridge. Test acceptance additionally checks the no-GCM control and a positive paired temporal-block interval for linear-minus-neural MSE. Bootstrap units are whole Mars-year/season blocks, not grid cells. The fixed budget is four aggregate GPU-hours, not four hours on each GPU.

Inspect scientific_success, test_evaluated, coverage and checksums. A stopped validation gate is a valid completed negative experiment. Synthetic tests do not establish held-out Mars skill; the full 176-start remote campaign is not validated by local tests.

Reproducibility

Keep native revision, source hashes, dependencies, terrain checksum, feature schema and physical settings with the run. Scientific changes need a new run directory. Completed forecasts are reused only if their contract matches.

This pipeline's trusted local pickle training checkpoints differ from the generic neural framework's pickle-free inference NPZ files. Do not load untrusted pickle artifacts. See the implementation recap.