btjd-data / README.md
NickGeissler's picture
Add license and links
d604bda verified
|
Raw History Blame Contribute Delete
2 kB
metadata
license: cc-by-4.0
pretty_name: BTJD datasets
tags:
  - stochastic-dynamics
  - generative-models
  - surrogate-modeling
  - turbulence

Block-Triangular Joint Drifting (BTJD): datasets

Datasets for the four experiments in Block-triangular joint drifting (BTJD): one-step generative surrogate models of stochastic dynamics (arXiv:2609.26435). Code, weights and the commands that reproduce every table and figure are at github.com/NickGeissler/btjd; scripts/download_data.sh there fetches this dataset into data/.

All arrays are NumPy .npy (or .npz), float32, indexed [trajectory, time, ...].

Path Shape Contents
duffing/paper/{train,test}.npy [5000, 1201, 2] Stochastic Duffing oscillator, X(0) ~ N([0,-10], I)
duffing/fixed_x0/test.npy [5000, 1201, 2] Same, from X(0) = (0, -10)
duffing/pool_{distIC,x0}/test.npy [200000, 1201, 2] Large pools for the trajectory QoI
rb/RB_train_mu{13.50,…,14.20}.npy [20000, 2001, 9] Stochastic Rayleigh–Bénard reduced model, 8 training μ
rb/RB_test_mu13.65.npy [20000, 2001, 9] Held-out μ = 13.65
burgers/StochBurgers{Train,Test}.npy [4096, 800, 64] Stochastically forced Burgers
cfd/stochasticJCFDTrain.npy [2048, 1000, 64, 64, 1] 2D stochastically forced turbulence, vorticity
cfd/stochasticJCFDTest.npy [1024, 1000, 64, 64, 1] Independent initial conditions, same forcing
cfd/test_normalized.npy [1024, 1000, 64, 64, 1] Test set divided by the training std (3.605973)
cfd/test_latent.npy [1024, 1000, 16, 16, 4] Test set encoded with the released autoencoder
cfd/operator_rollouts_ic0.npz runs: [3, 1000, 64, 64, 1] Operator-learning baseline rollouts from test IC 0

The Duffing and Rayleigh–Bénard sets can also be regenerated with duffing.generate_data / rb.generate_data in the code repository.