pflotran-run-workflow

Write perturbed input decks with write_parameter_file at each case's final location, then call create_case to stage static files (mesh, restart, database) around them. Validate with validate_adapter_ensemble.py --expect-baseline, then submit via submit_ensemble which sbatches each case's submit.sh.

Parse the *-mas.dat mass-balance columns and score through the target's own reduce function such as outflow_concentration or outflow_flux. Apply the 806-hour observation offset (measured hour 0 is model hour 806), and report per-species results beside any aggregate score.

The observation window carries an 806-hour offset: measured hour 0 corresponds to model hour 806. A window written as [0, 768] in model time contains zero observations and only spin-up. Check whether targets.yaml already applies the offset to avoid applying it twice.

No. The database file (e.g., savannah_river.dat) is a fixed input, not a calibrated surface. write_parameter_file raises NotImplementedError for surface="database", and the parameter list's EXCLUDED section explicitly designates the database as fixed input.

submit_ensemble returns synthetic DRYRUN-* job ids when A2MC_DRY_RUN is truthy or when sbatch is not on PATH. A dry run on a login node without SLURM looks identical to a real submission unless you inspect the job ids.

Serial execution is preferred: measurements showed 70.75% bit-identical values serially versus 68.49% with 8 ranks, and serial runs need only 32 MB and about one core-hour, letting roughly 128 members pack onto one node. A2MC_HPC_MPI_RANKS is set unconditionally in the site config for this reason.