Mechanism-Aware Ensemble Conditioning for Data-Limited Emulation of Extreme Events
Paper • 2609.30746 • Published
created_utc timestamp[s]date 2026-04-23 19:31:13 2026-04-23 19:48:17 | master_seed float64 6,647,249,545,382,099,000,000,000,000B 202,793,067,443,053,600,000,000,000,000B | semantics dict | cutoff dict | shapes dict | config dict |
|---|---|---|---|---|---|
2026-04-23T19:31:13 | 202,793,067,443,053,600,000,000,000,000,000,000,000 | {
"reference": "(n_runs, n_ref, 3)",
"coarse": "(n_runs, n_coarse, 3)",
"nudged": "(n_runs, n_coarse, 3)",
"ensemble": "(n_runs, ensemble_size, n_coarse, 3)",
"state_order": [
"z1",
"z2",
"z3"
],
"note": "Within each run, all trajectories share the same initial state; across runs, initial stat... | {
"cutoff_time": 0,
"ref_start_index": 0,
"coarse_start_index": 0,
"rule": "saved data starts at max(spinup_time, analysis_time)"
} | {
"time_reference": [
170000
],
"time_coarse": [
170000
],
"initial_states": [
2,
3
],
"reference": [
2,
170000,
3
],
"coarse": [
2,
170000,
3
],
"nudged": [
2,
170000,
3
],
"ensemble": [
2,
16,
170000,
3
]
} | {
"t_span": [
0,
510
],
"initial_state": [
0.1,
-1,
0.1
],
"n_ref": 170000,
"coarse_factor": 1,
"spinup_time": 0,
"tau": 8,
"n_runs": 2,
"ensemble_size": 16,
"sigma": 0.00001,
"analysis_time": 0,
"init_state_std": [
0.02,
0.02,
0.02
],
"a1": 2,
"a2": 2,
... |
2026-04-23T19:43:19 | 6,647,249,545,382,099,000,000,000,000,000,000,000 | {
"reference": "(n_runs, n_ref, 3)",
"coarse": "(n_runs, n_coarse, 3)",
"nudged": "(n_runs, n_coarse, 3)",
"ensemble": "(n_runs, ensemble_size, n_coarse, 3)",
"state_order": [
"z1",
"z2",
"z3"
],
"note": "Within each run, all trajectories share the same initial state; across runs, initial stat... | {
"cutoff_time": 0,
"ref_start_index": 0,
"coarse_start_index": 0,
"rule": "saved data starts at max(spinup_time, analysis_time)"
} | {
"time_reference": [
170000
],
"time_coarse": [
170000
],
"initial_states": [
2,
3
],
"reference": [
2,
170000,
3
],
"coarse": [
2,
170000,
3
],
"nudged": [
2,
170000,
3
],
"ensemble": [
2,
16,
170000,
3
]
} | {
"t_span": [
0,
510
],
"initial_state": [
0.1,
-1,
0.1
],
"n_ref": 170000,
"coarse_factor": 1,
"spinup_time": 0,
"tau": 8,
"n_runs": 2,
"ensemble_size": 16,
"sigma": 0.000001,
"analysis_time": 0,
"init_state_std": [
0.02,
0.02,
0.02
],
"a1": 2,
"a2": 2,
... |
2026-04-23T19:48:17 | 192,163,223,548,956,600,000,000,000,000,000,000,000 | {
"reference": "(n_runs, n_ref, 3)",
"coarse": "(n_runs, n_coarse, 3)",
"nudged": "(n_runs, n_coarse, 3)",
"ensemble": "(n_runs, ensemble_size, n_coarse, 3)",
"state_order": [
"z1",
"z2",
"z3"
],
"note": "Within each run, all trajectories share the same initial state; across runs, initial stat... | {
"cutoff_time": 0,
"ref_start_index": 0,
"coarse_start_index": 0,
"rule": "saved data starts at max(spinup_time, analysis_time)"
} | {
"time_reference": [
170000
],
"time_coarse": [
170000
],
"initial_states": [
2,
3
],
"reference": [
2,
170000,
3
],
"coarse": [
2,
170000,
3
],
"nudged": [
2,
170000,
3
],
"ensemble": [
2,
16,
170000,
3
]
} | {
"t_span": [
0,
510
],
"initial_state": [
0.1,
-1,
0.1
],
"n_ref": 170000,
"coarse_factor": 1,
"spinup_time": 0,
"tau": 8,
"n_runs": 2,
"ensemble_size": 16,
"sigma": 0.000005,
"analysis_time": 0,
"init_state_std": [
0.02,
0.02,
0.02
],
"a1": 2,
"a2": 2,
... |
ExtremeEventSensor
This dataset accompanies the NeurIPS 2026 paper Mechanism-Aware Ensemble Conditioning for Data-Limited Emulation of Extreme Events.
TLDR: Small ensembles of perturbed, low-cost simulations reveal locally unstable directions in the dynamics. We use that information to guide learned corrections to coarse trajectories.
What’s included
- LowDim: Trajectories from a three-dimensional nonlinear system exhibiting bursts, including reference, coarse, nudged, and perturbed ensemble simulations.
- QG: Data from a two-layer quasigeostrophic model, including upper and lower streamfunction with additional topography. Includes fields from reference, coarse, nudged, and ensemble simulations.
For a short demonstration, see the blog post. Full details about methods and examples are in the paper, and the accompanying code is on GitHub.
- Downloads last month
- -