Lessons¶
Apart from the REST_N default below, most setup problems were one-line script or namelist fixes.
Restarts and resubmission
Set REST_N explicitly. If it is left unset, it defaults to STOP_N. A 22-year segment killed at walltime then writes no restart. Use short segments (STOP_N=6) with REST_N no larger than STOP_N. Example: hillslope (Issues and fixes).
CESM3 restart pointers are dated. There is no undated rpointer.lnd; DRV_RESTART_POINTER must name the dated file, e.g. rpointer.cpl.2004-10-01-00000.
Keep NTASKS_ROF identical between spinup and evaluation. mizuRoute restarts record the task count, and a mismatch fails with “Number of MPI tasks on restart file is different from number using now”. Example: hillslope (Issues and fixes).
Start evaluations on a water year (RUN_STARTDATE=1980-10-01) so the record lines up with water-year analysis.
Namelist and build
History fields cannot change on a CONTINUE_RUN, because the field count must match the restart. Add fields such as H2OSOI at the start of a new (hybrid) run.
Case-specific lessons
-irrig on rejected: irrigation_on (Issues and fixes).
Build-namelist defaults with flanduse_timeseries, and regional grids: evolving_landcover (Issues and fixes).
QLATFLOWOUT and hillslope with use_init_interp (Issues and fixes): hillslope.
Diagnosis
Check the forcing before blaming the model. Example: control (Spinup).
Check seasonal signals before calling a restart bug. Example: hillslope (Spinup).
Caveats¶
The streamflow skill numbers stand as produced, but four known issues limit how far differences can be attributed to land processes.
No NLDAS2-specific parameter tuning. See Common setup (Parameter tuning row of the shared setup table).
The land-cover run also has irrigation on. See evolving_landcover (case card, Baseline).
Spinup differs by configuration. See hillslope (Spinup), control (Spinup) and prog_phenology (Spinup).
Gauge types are mixed. Some gauges report naturalized flows and others observed (regulated) flows; skill is not yet split by type.
Near-zero-flow gauges: arid gauges with near-zero flow can have extreme percent bias (North Concho River, TX: +16,293%) and are treated as unreliable tests of skill; this is a working judgment, not a tested criterion.
Scope not yet covered: snow, soil moisture and ET have not been evaluated against NLDAS or independent observations. Results are monthly only, because mizuRoute daily output was not written.
Future work¶
With more time and funding, the next steps are a full non-streamflow evaluation, cleaner single-option contrasts, and publishing the runs as baseline datasets. These are candidates for a follow-on proposal, beyond the SIF-T scope; none is decided.
Evaluation
Evaluate snow, soil moisture and ET against NLDAS and independent observations. The SIF-T application promised this, and it was not started.
A gridded-layer dashboard: CONUS maps of SWE, QRUNOFF, soil moisture, ET, precipitation, T2M and LH/SH for each configuration and as differences from its baseline. Andy Wood asked for these fields; a version was tested on synthetic data only.
Skill split by naturalized vs observed gauges and by water yield, with a gauge-type filter in the dashboard.
Experiments
Rerun evolving land cover with irrigation off, for a clean land-cover contrast against control.
Irrigation with limit_irrigation_if_rof_enabled = .true. (and optionally groundwater irrigation), so irrigated runs conserve basin water.
A longer hillslope spinup, to remove the residual soil-moisture drift.
A full carbon spinup for prog_phenology, so carbon fluxes can be evaluated too.
Parameter tuning or calibration for NLDAS2 forcing.
Products
Publish the five simulations as community baseline datasets with a DOI.
Machine-learning emulators trained on the best-performing configurations.