Skip to content

CHANNEL_CONVERGENCE_PLAN.md

Latest commit

 

History

History
73 lines (59 loc) · 3.88 KB

CHANNEL_CONVERGENCE_PLAN.md

File metadata and controls

73 lines (59 loc) · 3.88 KB
Raw

Channel side-plan: LINEAR-EOS truth + resolution convergence (2026-06-25)

Purpose. (1) Re-create the deleted LINEAR woc_p0625 truth, and (2) run one finer (p03125) wind to demonstrate that the large-scale metrics are converged at p0625. Both staged as warm-starts (change one thing per run). Sits inside the bigger channel-run program for Part 2 §4 (see PLAN.md "CHANNEL-RUN PLANNING").

Config. extra_sponge_slow + woc (LINEAR EOS, density-space diagnostics) — the analysis-grade, theory-matched family. Headline wind τ=0.2 (single wind for both stages).

Term note. "warm-start" = restart from the previous equilibrium, not from rest.

  • Stage 1 = true restart-from-file (Z* prognostics u,v,h,T,S carry over; density is diagnosed, so flipping EQN_OF_STATE needs no interpolation — the state re-adjusts).
  • Stage 2 = re-initialize from interpolated fields (different grid).

Stage 0 — prep (VERIFIED 2026-06-25)

  • LINEAR woc override block (from surviving woc coarse runs): EQN_OF_STATE=LINEAR, LIGHTEST_DENSITY=1026.2, NUM_DIAG_COORDS=2 (z+rho2), time-mean diag_table (prog_tmean/prog_z_tmean/prog_rho_tmean). Also enable CALCULATE_APE=True (APE in ocean.stats — free; its use is under discussion but output is harmless) and register PE_to_KE (3D) in the diag_table (energy-conversion metric).
  • Sponge — ✅ CLEARED (was the main risk): sponge restores potential temperature (SPONGE_PTEMP_VAR="Temp"), so it is EOS-independent → reuse sponge_damping_NI640_NJ368_relax_360_days.nc as-is for LINEAR (no regen). Stage 2 still needs a NEW ..._NI1280_NJ736_... sponge — but since the target is a T field, just regrid/interpolate it.
  • Restart — ✅ verified: WRIGHT p0625 tau_0.2 RESTART/ has full MOM.res.nc (749 MB) + ocean_solo.res at the equilibrated day-~10000 end-state → clean restart-from-file.

Stage 1 — EOS switch @ p0625 (WRIGHT → LINEAR)

  • Restart existing WRIGHT extra_sponge_slow_p0625/tau_0.2_... from its restart file with the LINEAR woc block. No interpolation.
  • Run to re-equilibration (monitor ACC transport + domain KE); expect ~8–10 model-yr (slowest adjuster = 360-day sponge / deep stratification).
  • ~5 h wall on 480 cores (6 nodes), 1 job.
  • Output: LINEAR woc_p0625 truth (τ=0.2) — doubles as launch state for Stage 2.

Stage 2 — resolution jump @ p03125 (LINEAR)

  • Interpolate equilibrated LINEAR p0625 → p03125 (1280×736×30) as IC. DT 900→450 s, layout ~48×40, ~24 nodes (4× memory).
  • Run ~6 model-yr (1–2 yr fine-eddy spin-up + ~3–4 yr averaging).
  • ~6–8 h wall, 1–2 chained 16 h jobs.
  • Output: LINEAR woc_p03125 (τ=0.2).

Stage 3 — convergence analysis

  • Compare p03125 vs p0625 (both LINEAR, τ=0.2): ACC transport, Ψ(y,ρ) max + structure, isotherm/thermocline depth, domain EKE.
  • Adaptive stop: small delta → converged, write it up; large delta → then consider p015625.

Cost

stage model-yr cores wall
1 (EOS switch p0625) ~8–10 480 (6 nodes) ~5 h
2 (p03125 convergence) ~6 ~1920 (24 nodes) ~6–8 h

Baseline measured: p0625 = 27.4 model-yr in 13.3 h on 480 cores (≈233 core-h/model-yr); p03125 ≈ 8× compute, 4× memory.

Implementation deltas for compute-Claude (vs existing launcher)

  • Restart-from-existing-run path (Stage 1) instead of cold base_exp.
  • p03125 resolution preset (NI1280×NJ736) + its sponge file.
  • p0625→p03125 interpolation of the IC.

Open decisions

  1. Headline wind — τ=0.2 assumed.
  2. Stage-1 truth breadth — τ=0.2 only now; extend to more winds later if the online-eval truth needs them (don't over-build before convergence result is in).

Status: plan written 2026-06-25; not yet launched. Awaiting headline-wind confirm + big-picture run-matrix decisions before handing the spec to compute-Claude.