README.md

---
pretty_name: ChaosNetBench-CML
license: cc-by-4.0
language:
  - en
tags:
  - chaos
  - dynamical-systems
  - spatio-temporal-forecasting
  - graph-neural-networks
  - physics
configs:
  - config_name: default
    data_files:
      - split: results_preview
        path: data/multiseed_aggregated.csv
---

# Dataset Card - ChaosNetBench-CML

Primary dataset file: [data/chaosnetbench_cml.h5](https://huggingface.co/datasets/htmoges/chaosnetbench-cml/resolve/main/data/chaosnetbench_cml.h5)

## Release

- Version: 1.0.0
- Public release date: 2026-05-09
- Creator and maintainer: H. T. Moges
- Corresponding contact: ht.moges@gmail.com
- Homepage: https://htmoges.github.io
- Code: https://github.com/htmoges/ChaosNetBench
- License: CC-BY 4.0

## Purpose

ChaosNetBench-CML is a benchmark dataset and evaluation framework for systematically comparing spatio-temporal graph neural networks on controlled chaotic lattice dynamics. Built on coupled standard maps with known ring topology and independently tunable local chaos (`K`), coupling (`epsilon`), and system size (`N`), it supports regime-aware comparisons between graph-aware and purely temporal baselines across 96 system instances and 9,600 trajectories.

## Contents

| Asset | Description | Size |
|---|---|---|
| Full HDF5 dataset (`data/chaosnetbench_cml.h5`) | All 96 system instances with 100 initial conditions each, including `state_wrapped`, initial conditions, and SALI-based orbit labels | ~27.3 GB |
| `data/multiseed_aggregated.csv` | Lightweight public preview of aggregated benchmark results | ~350 KB |
| `data/results_preview/README.md` | Provenance note for the benchmark-results preview | ~1 KB |
| `metadata/chaosnetbench_cml.croissant.json` | Full-dataset Croissant metadata | ~10 KB |
| `metadata/chaosnetbench_cml_results_preview.croissant.json` | Croissant metadata for the results preview | ~8 KB |

### CSV schema

| Column | Description |
|---|---|
| `model` | Model identifier string |
| `K` | Standard map nonlinearity in {0.5, 0.97, 2.0, 6.5} |
| `rho` | Coupling ratio rho = epsilon / K in {0.05, 0.075, 0.10, 0.15, 0.20, 0.30, 0.40, 0.50} |
| `N` | Number of oscillator sites in {8, 16, 32} |
| `test_mse_mean` | 3-seed mean test-window MSE |
| `test_mse_std` | Across-seed standard deviation of test MSE |
| `ar_vpt_mean` | 3-seed mean autoregressive Valid Prediction Time |
| `ar_vpt_std` | Across-seed standard deviation of AR VPT |

## Metric definitions

Metric definitions are implemented in `chaosnetbench/metrics.py` in the public code repository:

- VPT (Valid Prediction Time): first autoregressive step where NRMSE > 1.0. NRMSE is RMSE normalized by the standard deviation of the true trajectory.
- Convergence filter: runs with `test_mse_mean >= 0.95` are degenerate (near-constant output) and are excluded from VPT head-to-head comparisons.
- 3-seed aggregation: each `(model, K, rho, N)` configuration is trained with 3 random seeds; `ar_vpt_mean` and `test_mse_mean` are the means, while `*_std` fields are across-seed standard deviations.

## Generation protocol

System: Coupled Standard Map (Chirikov-Taylor map, ring topology, nearest-neighbor coupling).

Parameter grid:
- `K` in {0.5, 0.97, 2.0, 6.5}
- `rho` in {0.05, 0.075, 0.10, 0.15, 0.20, 0.30, 0.40, 0.50}; `epsilon = rho * K`
- `N` in {8, 16, 32}

Protocol:
- 100 initial conditions per configuration (70 train / 10 validation / 20 test)
- IC-based split to avoid temporal leakage
- 1000 transient steps discarded + 10000 recorded steps
- SALI orbit classification with 1000 tangent-map iterations and early termination at `SALI < 1e-8`

Code for trajectory generation and benchmark evaluation is available in the public repository: https://github.com/htmoges/ChaosNetBench

## Results preview

The lightweight `multiseed_aggregated.csv` preview was produced by:
1. Training each benchmark model on the full trajectory dataset for 50 epochs with 3 random seeds.
2. Evaluating on the IC-held-out test split with 20 initial conditions per configuration.
3. Computing per-seed VPT and test MSE.
4. Aggregating the results into one CSV row per `(model, K, rho, N)` with cross-seed means and standard deviations.

The preview contains post-aggregation metrics only. It is intended to make the benchmark outputs easy to inspect without downloading the full HDF5 release or retraining models.

## Responsible AI

- Data limitations: covers only coupled standard map lattice dynamics; results may not generalize to dissipative systems, continuous-time chaotic flows, or spatiotemporal PDEs.
- Data biases: entirely synthetic; no human subjects. The parameter grid intentionally samples critical transition regimes for scientific analysis.
- Personal or sensitive information: none.
- Validated uses: model comparison under controlled chaos; STGNN-versus-temporal baseline evaluation; analysis of rollout robustness.
- Synthetic data: yes.