{"success":true,"data":{"author":"Jorge (jorge-mif)","bucketAccessType":"public","bucketName":"jorge-mif/kolmogorov","bucketPrefix":"","category":"Physics","cloudProvider":"huggingface","createdAt":"2026-11-21T20:46:01.274Z","description":"Simulated trajectories of deterministically forced 2D incompressible Navier-Stokes flow (Kolmogorov flow) in vorticity form with forcing sin(4y) and linear drag 0.1, generated with the JAX-CFD spectral solver (Crank-Nicolson RK4, anti-aliased) on a periodic 256 x 256 grid with viscosity 1e-4 (Reynolds number about 10^4). The repository holds 64 independent initial conditions with 200 saved frames each as float32 NumPy arrays of shape (64, 200, 256, 256). The spacing between saved frames (0.1 time units, 100 solver steps of dt = 1e-3) is set to about one sixth of the measured Lyapunov time (leading Lyapunov exponent about 1.56 from twin-trajectory renormalization), so consecutive states stay correlated while chaotic divergence spreads the conditional distribution; a 5,000-step warm-up is discarded so frames sample the statistically stationary regime. Intended for probabilistic and generative forecasting of chaotic dynamics, stochastic interpolants and diffusion-model PDE surrogates. Load with numpy.load (memory-mapped). Released under CC BY 4.0 by the author.","displayTitle":"Kolmogorov Flow at Re ~10^4, Lyapunov-Calibrated Trajectories for Probabilistic Neural PDE Forecasting (JAX-CFD)","huggingFaceUrl":"https://huggingface.co/datasets/jorge-mif/kolmogorov","id":"c18e7e14-df15-5aa0-8ac8-c46fee6bd74a","license":"CC-BY-4.0","licenseEvidenceUrl":"https://huggingface.co/datasets/jorge-mif/kolmogorov","objectCountEstimate":5,"probeStatus":"ok","region":"us","schemaCard":{"accessHints":[{"language":"python-datasets","snippet":"from datasets import load_dataset\nds = load_dataset(\"jorge-mif/kolmogorov\")"},{"language":"python-hub","snippet":"from huggingface_hub import snapshot_download\nsnapshot_download(repo_id=\"jorge-mif/kolmogorov\", repo_type=\"dataset\")"}],"caveats":["HuggingFace metadata: author=jorge-mif, license:cc-by-4.0, size_categories:10K<n<100K, downloads=66"],"format":"mixed","generatedAt":"2026-11-21T20:48:52.674Z","layout":{"example":"kolmogorov_Re1000_64_256x256_200steps_fullres.npy","pattern":"{repoId}/{filename}"},"mountPath":"/mnt/archive","sampleHeaders":[{"bytes":"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","decoded":{"fields":{"descr":"<f4","dtype":"float32","fortranOrder":false,"ndim":4,"shape":"64,200,256,256"},"format":"npy","summary":"NPY float32 [64, 200, 256, 256]"},"encoding":"hex","key":"kolmogorov_Re1000_64_256x256_200steps_fullres.npy","truncatedAt":512}],"sampleObjects":["kolmogorov_Re1000_64_256x256_200steps_fullres.npy","kolmogorov_Re1000_64_256x256_200steps_fullres_indices.npy","kolmogorov_Re1000_64_256x256_200steps_std.npy"],"sidecars":[{"bytes":"---\nlicense: cc-by-4.0\ntags:\n- physics\n- turbulence\n- pde\n- navier-stokes\n- chaotic-dynamics\n- forecasting\nsize_categories:\n- 10K<n<100K\npretty_name: kolmogorov\n---\n\n# Kolmogorov Flow (Lyapunov-calibrated)\n\nDeterministically forced 2D Navier–Stokes (Kolmogorov flow) at \\\\(Re = 10^4\\\\), sampled at a forecast step calibrated to the system's measured Lyapunov time. Intended for probabilistic / generative neural PDE forecasting on chaotic dynamics. \n\nGenerated with [JAX-CFD](https://github.com/google/jax-cfd)\n\n## Summary\n\n- **Equation.** Vorticity formulation of 2D incompressible Navier–Stokes with deterministic Kolmogorov forcing  \\\\(f = \\sin(k_f y)\\,\\hat x \\\\) and linear drag.\n- **Solver.** Spectral (jax-cfd, `crank_nicolson_rk4`), anti-aliased.\n- **Resolution.**  \\\\(256 \\times 256 \\\\), periodic on  \\\\([0, 2\\pi]^2 \\\\).\n- **Viscosity.**  \\\\(\\nu = 10^{-4} \\\\) (Re  \\\\(\\sim 10^4 \\\\).\n- **Trajectories.** 64 independent ICs, 200 saved frames each.\n- **Format.** Single `.npy`, shape `(64, 200, 256, 256)`, `float32`.\n\n## Time-step calibration\n\nThe forecast step  \\\\(\\Delta\\tau \\\\) between saved frames is set as a fraction of the system's leading Lyapunov time so that consecutive states are dynamically correlated but small initial differences have grown to a macroscopic fraction of the field.\n\nMeasured Lyapunov exponent (twin-trajectory, periodic renormalization, after spin-up):\n$$\\Lambda \\approx 1.56\\;\\;\\text{(time}^{-1}\\text{)}, \\qquad T_\\Lambda = 1/\\Lambda \\approx 0.64.$$\n\nSolver `dt = 1e-3`, `INNER_STEPS = 100` (solver steps between saved frames), giving:\n$$\\Delta\\tau = 100 \\times 10^{-3} = 0.1, \\qquad \\Delta\\tau / T_\\Lambda \\approx 0.16.$$\n\nSo consecutive saved frames are separated by roughly one-sixth of a Lyapunov time. Two states that share a coarse-resolution projection but differ in unresolved continuum directions have time to diverge by a factor  \\\\(e^{0.16} \\approx 1.17 \\\\) between frames.\n\nA warm-up of 5000 solver steps ( \\\\(\\sim 5\\,\\tau_{\\text{eddy}} \\\\) is discarded per trajectory before saving, so frames are samples of the statistically stationary regime.\n\n## Why this calibration\n\nProbabilistic forecasters that model  \\\\(p(u_{n+1} \\mid u_n) \\\\) as a stochastic interpolant need the conditional to have non-trivial width on the discrete state. If  \\\\(\\Delta\\tau \\ll T_\\Lambda \\\\), similar  \\\\(u_n \\\\) are followed by nearly identical  \\\\(u_{n+1}\\\\) and the conditional collapses to a delta — there is no distribution to learn. The choice  \\\\(\\Delta\\tau \\sim 0.1\\text{–}0.25\\,T_\\Lambda \\\\) is the smallest horizon at which chaotic divergence produces a genuinely spread conditional while consecutive frames remain dynamically related (i.e. it is still forecasting, not decorrelated sampling).\n\n## Physical setup\n\n| Parameter | Value |\n|---|---|\n| Domain |  \\\\([0, 2\\pi]^2 \\\\), periodic |\n| Grid |  \\\\(256 \\times 256 \\\\) |\n| Viscosity  \\\\(\\nu \\\\) |  \\\\(10^{-4} \\\\) |\n| Reynolds |  \\\\(\\sim 10^4 \\\\) |\n| Forcing |  \\\\(f = \\sin(4y)\\,\\hat x \\\\) (deterministic) |\n| Drag |  \\\\(-0.1 \\\\) |\n| Solver | jax-cfd spectral, CN-RK4 |\n| `dt` |  \\\\(10^{-3} \\\\) |\n| Inner steps / frame | 100 |\n| \\\\(\\Delta\\tau\\\\) |  \\\\(0.1 \\\\) |\n| \\\\(T_\\Lambda\\\\) (measured) |  \\\\(\\approx 0.64\\\\) |\n| Warm-up | 5000 solver steps |\n| Trajectories | 64 |\n| Frames / trajectory | 200 |\n\n## Loading\n\n```python\nimport numpy as np\ndata = np.load(\"kolmogorov_Re1e4_64x200x256x256.npy\")  # (64, 200, 256, 256), float32\n```\n\nTrajectories are independent (different IC seeds, same physics). Common split: first $N$ trajectories train, remainder validation.\n\n## Citation\n\nIf you use this dataset, please cite both the underlying solver:\n\n```\n@article{Kochkov2021-ML-CFD,\n  author = {Kochkov, Dmitrii and Smith, Jamie A. and Alieva, Ayya and Wang, Qing and Brenner, Michael P. and Hoyer, Stephan},\n  title = {Machine learning{\\textendash}accelerated computational fluid dynamics},\n  volume = {118},\n  number = {21},\n  elocation-id = {e2101784118},\n  year = {2021},\n  doi = {10.1073/pnas.2101784118},\n  publisher = {National Academy of Sciences},\n  issn = {0027-8424},\n  URL = {https://www.pnas.","encoding":"utf8","key":"README.md","name":"README.md","size":4850,"truncated":true}],"sourceObjectKey":"kolmogorov_Re1000_64_256x256_200steps_fullres.npy","starterCell":"from huggingface_hub import hf_hub_download\npath = hf_hub_download(repo_id=\"jorge-mif/kolmogorov\", filename=\"kolmogorov_Re1000_64_256x256_200steps_fullres.npy\", repo_type=\"dataset\")\nprint(path)","starterSamplePath":"kolmogorov_Re1000_64_256x256_200steps_fullres.npy","totalObjects":5},"summary":"Simulated trajectories of deterministically forced 2D incompressible Navier-Stokes flow (Kolmogorov flow) in vorticity form with forcing sin(4y) and linear drag 0.1, generated…","title":"Kolmogorov Flow at Re ~10^4, Lyapunov-Calibrated Trajectories for Probabilistic Neural PDE Forecasting (JAX-CFD)","updatedAt":"2026-11-21T22:17:21.669Z","visibility":"public","websiteUrl":"https://github.com/google/jax-cfd"}}