The Dataset Viewer has been disabled on this dataset.

samudrace-circuits: data products

Derived data for the interpretability study of SamudrACE C1A0 (HF mahf708/e3smv3-lr-hist-samudrace@rc1) against E3SMv3 truth (v3.LR.historical_0101.aigo). Code and docs: https://github.com/mahf708/samudrace-circuits (start with HACKATHON.md). Results page: https://portal.nersc.gov/project/m5451/mahf708/samudrace-circuits/

Path What
states/t1995, states/t2045 C1A0 coupled runs started from E3SM truth at 1995-01-01 / 2045-01-01, 73 Γ— 5-day steps: every ocean network input x_###.npy (normalized, land = 0), meta.json/meta.npz (names, normalizer, mask), fme ocean predictions and truth targets (fme/ocean/), daily atmosphere spectra (atm_spectra.npz)
states/ic1985 the laptop-reproduction run from the 1985 release IC (a training year): inputs x_### and outputs y_###
states/atm12, states/atm12m5 first atmosphere network call (input atm_x_00.npy, output atm_y_00.npy; batch of 12) at Jan 16 / Apr 16 / Jul 15 / Oct 13 of 1996, 2042, 2047 (and those dates βˆ’ 5 days), plus the first ocean inputs (atm12)
f1/year_YYYY.npz per 5-day window (73/yr): ocean sst, the five surface heat fluxes as the training loader signs them, ice, and EAM TS at window start/end, window-mean TS⁴ and EAM FLUS; 1985–89, 1995–99, 2040–49
rollout1y/t2040 daily atmosphere spectra of a 1-year run from 2040-01-01
results/ all result JSONs and npz (slope maps, truth conditional maps, gain spectra)

Train years 1940–1989, 2000–2039; validation 1990–1994; never seen 1995–1999, 2040+. Not an Ai2 release.

Downloads last month
-