This file maps each figure/table of "Finding the potential height of tropical cyclone storm surges in a changing climate using Bayesian optimization" (EarthArXiv https://doi.org/10.31223/X57T5R, under revision at Environmental Data Science) to the command that produces it, its key inputs, and its output path.

Note on years (updated 2026-07-05): the revised paper's BO results are the 2015 vs 2100 experiments (CESM2 r4i1p1f1, SSP5-8.5 August monthly means) throughout — matching the repository's experiment directories, figure filenames, and slurm variables. The original submission's 2025/2097 runs (r10i1p1f1 single-site) are retired; the legacy 2025.json/2097.json profiles in w22/data are kept for the record only (NOTE: they store p in Pa, unlike the hPa convention of the shipped city profiles — adforce.profile.read_profile now rejects them loudly).

Known issue — potential-size humidity fix (2026-07-07). The Wang-2022 y -> p_m back-conversion in w22/ps.py dropped the ambient relative-humidity factor from its numerator (all three solver entry points), so solved potential sizes were effectively the rh = 1 answer regardless of the input humidity (canonical point: r0 2136.9 -> 2079.5 km at rh = 0.9, i.e. -2.7%; rmax -4.5%). Fixed via the shared w22.w22_carnot.carnot_pm_from_y (see tests/test_ps_units.py; a second silently-ignored input key, cd_ck vs ck_cd, was fixed at the same time — golden pins now sit 2.9% under Wang 2022's canonical r0). Artifacts computed BEFORE the fix and not yet regenerated:

• the six w22/data/*_profile_r4i1p1f1.json city profiles and the PI/PS input table (thesis worstpp:tab:pips r3/ra3 columns). At the CMIP6 surface relative humidities feeding these points (August Gulf/Florida, typically ~0.75-0.85, vs the effectively-rh=1 answer the buggy solver returned), the rh sweep implies the solved sizes shrink roughly 4-8% (canonical-point sweep: rh 0.8 -> -6.0%, 0.9 -> -2.9%, 0.95 -> -1.6%);
• ALL gridded PS products: the ERA5 PS zarr (tcpips/era5.py::era5_calculate_ps), the CMIP6 PS products (tcpips/ps_driver.py, w22/ps_runs.py) — and therefore the paper's PS map/time-series panels (Fig. 2), the ~2.3 km/yr potential-size trend and the SST-PS regression statistics;
• the ERA5/IBTrACS along-track PS validation outputs (tcpips/ibtracs.py);
• the tradeoff/4D BO pathway, contaminated DIRECTLY (w22/tradeoff.py calls calculate_ps_ufunc live per sample, bypassing the profile JSONs) — any pre-fix tradeoff/4D runs, and data/adbo/gpr_data.csv, whose surge values come from ADCIRC runs forced by the pre-fix New Orleans profile (the GP kernel RANKING is plausibly robust, the absolute skill numbers are pre-fix);
• the thesis figure-4a/4b artifacts (img/w22/figure_4a.pdf, figure_4b.pdf and the solver-output CSVs 4a-*-ours.csv, 4b-*-ours.csv in w22/data/w22/) — still pre-fix builds; regenerate via python -m w22.test_figures / W22_SLOW_TESTS=1 pytest w22/test_figures.py;
• (downstream, via the profiles) the ADCIRC potential-height experiments.

Regenerating the profiles requires the CMIP6 processed PI/PS point inputs (data/cmip6/regrid/... AND the potential-intensity products data/cmip6/pi*, ARCHER2 backup); regenerating potential heights requires ADCIRC re-runs on ARCHER2. SST and potential-intensity (Vp) values are NOT affected (solver inputs, computed by tcpyPI upstream of the PS solve).

See the README for environment setup (env.yml portable spec; environment.lock.yml/requirements.lock.txt exact validated solve) and pip install -e .[bo] etc. for the optional dependency groups needed by each stage. slurm/README.md documents the ARCHER2 job scripts referenced below.

The end-to-end chain is:

CMIP6/ERA5 download (tcpips) → regrid (tcpips) → potential intensity (tcpips.pi) → potential size (w22.ps) → ADCIRC forcing (adforce) → Bayesian optimization (adbo.exp*) → EVT with an upper bound (worst)

Stages 6–7 require ARCHER2 (or another HPC machine with a ported adforce/wrap.py and compiled ADCIRC ≥ 55.02); everything else can run locally. The archived outputs of stages 1–7 shipped in data/ and the Hugging Face datasets let the figure scripts below run without re-doing the HPC stages.

Some plotting entry points are driven from commented-out calls in the module __main__ blocks (noted below); uncomment the relevant call and run python -m <module>.

These are produced by the repository and appear in the thesis or support the analysis, but have no figure/table number in worst_new.tex: