Argo-like 3D observations and ocean DA problem redesign

[glwagner]

Summary

• Replace surface-only observations with Argo-like 3D profile sampling in the ETKF: 76 profiles/cycle, 31 vertical levels in upper 1000m, ~4,700 total obs (T+S) -- down from 56K surface-only
• Add daily 3D tracer output to nature run for subsurface observation extraction
• Add RESEARCH_NOTES.md documenting literature review on ocean DA for hurricane prediction

Key finding: our current problem is too easy

Surface perturbations (0.3 C bias + 0.5 C noise) decay naturally under atmospheric forcing in ~11-14 days. The 32% RMSE reduction we observed is mostly the atmosphere doing the work, not the DA.

The scientifically meaningful problem is recovering subsurface structure (thermocline depth, ocean heat content at 50-300m), which:

• Persists for months-years without DA
• Is poorly observed (Argo only, every 10 days)
• Directly controls hurricane intensity (OHC > 60 kJ/cm2 supports intensification)

Changes

Argo-like observation network (ensemble_da.jl)

• 76 random profile locations per cycle (Argo density: ~1 per 3-degree box, profiling every ~10 days)
• Each profile observes all 31 grid levels in upper 1000m (k=20:50)
• Total obs (T+S): ~4,700 per cycle (was 56K surface-only)
• Obs-to-ensemble ratio: 470:1 (10 members), 47:1 (100 members)
• MPI-aware: local extraction + single Allreduce for full ensemble-obs matrix
• Land profiles filtered using nature run surface T mask

Nature run 3D output (nature_run.jl)

• Daily 3D T/S snapshots via JLD2Writer (TimeInterval 1 day)
• ~5 GB total for 14-day run (very manageable)
• Required for subsurface observation extraction

Research notes (RESEARCH_NOTES.md)

• Literature review on ocean DA for hurricane prediction
• Operational DA window lengths (3-7 days for 4DVAR)
• Perturbation persistence analysis: surface decays in days, subsurface persists months-years
• Recommended redesign: subsurface perturbations, multi-platform obs, hurricane-relevant metrics

Next steps

1. Subsurface perturbations: zero at surface, max ~1.5 C at 150m, spatially correlated (~200 km)
2. Multi-platform obs: Argo profiles + satellite SST + SSH altimetry
3. Localization (via Manteia.jl) to prevent ensemble collapse
4. Scale to 100 ensemble members
5. Hurricane-relevant metrics: OHC, mixed layer depth, thermocline depth

Test plan

• Re-run nature run with 3D output (SLURM currently in maintenance)
• Run ETKF with Argo-like observations
• Verify land masking works (profiles on land filtered out)
• Compare spread/RMSE evolution with previous surface-only run

Generated with Claude Code