forecasting-models / code /configs /sqg /dataset.yaml
coarse-grained-forecasting-benchmark's picture
Add anonymous reproduction code
41dd7e0 verified
Raw History Blame Contribute Delete
1.36 kB
# Surface quasi-geostrophy (pyqg): the package `sqg_v3_candidate_v2`. Set SQG_DATA and SQG_TRUTH_BANK.
#
name: sqg_v3_candidate_v2
system: sqg
data: ${SQG_DATA}
truth_bank: ${SQG_TRUTH_BANK}
expected_sha256: {trajectories_fine.npz: 510b027125b30065489edbae1d0ac23c3cb15765630be4113c119af993c0dee4, split.json: fb482ffd0353ba47dfca67a34466f7026278962ca9c8fb44839aa039935d08e2,
normalization.npz: 46f94fc794069c9ad985d23ae4cc2b8ed866fe13d6a0a5463f1c8384352898f8}
state:
dim: 196
observed: coarse spectral surface buoyancy b, |k| <= 8, Hermitian half (98 complex modes), b_hat / N^2
hidden: a Gaussian buoyancy field on 8 < |k| <= 32 with |b_k|^2 ~ k^-5/3 and rms 0.1, drawn per realization
physics: {solver: pyqg.SQGModel, nx: 256, dt: 0.005, L: 2*pi, H: 1.0, Nb: 1.0, f_0: 1.0, beta: 0.0, rek: 0.0, T_phys: 20.0}
time_grid: {n_snapshots: 41, dt: 0.025, T: 1.0, note: benchmark time 1.0 = 20 SQG time units}
package: {vecs_train: [80, 200, 41, 196], vecs_eval: [20, 200, 41, 196]}
splits:
train: 80 conditions x 200 realizations
val: 10 conditions x 200 realizations
test: conditions 100-109, 200 realizations in the package; 10,000 each in the truth bank
normalization: per-coordinate mean and std of the 80 training conditions x realizations j < 20 x all 41 snapshots
fid_adapter: {name: passive_scalar_coarse, params: {n_disp: 128, k_c: 8}}