ydai17's picture
Add files using upload-large-folder tool
7eeac21 verified
Raw
History Blame Contribute Delete
1.81 kB
{
"CAVEATS": [
"Stored values are RAW physical fp32 fields (no normalization applied). Load solver.normalizer.FrozenNormalizer.from_json(decay_saffman_v2_norm_minmax.json or decay_saffman_v2_norm_standardize.json) to transform at load time; velocity (u,v,w) is normalized JOINTLY (one shared scale) to preserve A10 component isotropy. NOTE: the 'standardize' mode is UNBOUNDED (z-score, ~+/-5, pressure can reach ~-11); only 'minmax' maps to [-1,1].",
"Normalization constants were fit on seeds [2, 3, 7] (see attrs['norm_fit_scope']). For train-only fits, held-out frames MAY slightly exceed [-1,1] under 'minmax' -- that is correct (the scale is the TRAIN absmax), not a bug.",
"Frames are filtered by instantaneous k_maxeta>=1.5 and are NOT uniformly spaced in time -- use the stored 't' array (monotonic WITHIN each seed, resets at seed boundaries), do not assume uniform dt.",
"The 'seed' array stores the TRUE IC pool number (_corpus->0, _corpus_poolK->K), so a loader can filter the zarr by (case, seed) exactly as the by-IC split manifest emits; rejected seeds are simply absent (no compaction/relabeling).",
"The corpus under-samples the strong-dissipation (eps-peak) tail; high-order intermittency statistics are a lower bound (see DATASHEET.md)."
],
"case": "decay_saffman_v2",
"channels": [
"velocity (u,v,w joint)",
"p"
],
"complete": true,
"compute_precision": "fp64",
"n_frames": 400,
"n_seeds": 8,
"norm_fit_scope": "decay_saffman_v2: fit on 150/400 frames from seeds [2, 3, 7] (--seed-filter train; 20260715T104201Z)",
"normalizer": "solver.normalizer.FrozenNormalizer (velocity JOINT, vetted q2)",
"pressure_source": "operators.pressure_hat (D4-certified)",
"storage_precision": "fp32"
}