phase0-design

Create a per-round config wrapper sourcing the base site config, design a new parameter-list CSV with documented bounds, then sample with create_adapter_parameter_sample.py, materialize cases with materialize_adapter_ensemble.py, validate, and submit. The skill walks through each step including the required parameter-set design review.

Four methods are supported: Morris (trajectory-based screening), Saltelli Sobol (variance decomposition), LHS, and sobol_seq (a scrambled low-discrepancy sequence for surrogate training). The FATES path offers the first three; the adapter path adds sobol_seq, which has no SALib analyzer.

Use submit_adapter_ensemble_batched.py, which submits in queue-aware waves against the QOSMaxSubmitJobPerUserLimit with idempotent resume via job_id.txt. Alternatives include a SLURM job array template or a packed node-exclusive script when the shared partition is the bottleneck.

Yes, adapter models like EcoSIM, PFLOTRAN, and ATS use parallel scripts that dispatch through a ModelBackend instead of the FATES-specific tools. PFLOTRAN has its own validation script because its parameter ids and text input deck differ from NetCDF-based models.

A dropped parameter does not fall back to the CSV default column; its value flows from the base parameter file into every case. Always read the value directly from the base file for each dropped parameter and record it in the round log.

Use sobol_seq when training a surrogate model, since its exactly-N space-filling points avoid the optimistic cross-validation bias of Saltelli's cross-sampled design. Use Saltelli Sobol when estimating variance-based sensitivity indices.