ChaosNetBench

ChaosNetBench: Benchmarking Spatio-Temporal Graph Neural Networks on Chaotic Lattice Dynamics

Official codebase and public benchmark interface for ChaosNetBench.

ChaosNetBench is a controlled benchmark for studying spatio-temporal graph neural networks (STGNNs) on the lattice of coupled standard maps with known topology, tunable chaotic regimes, and initial condition based evaluation.

Links

• Paper: https://arxiv.org/abs/2605.09676
• Dataset: https://huggingface.co/datasets/htmoges/chaosnetbench-cml
• Maintainer: H. T. Moges
• Contact: ht.moges@gmail.com
• Homepage: https://htmoges.github.io

Physical system: The Coupled Standard Map (Chirikov-Taylor map on a ring lattice) was introduced as the benchmark dynamical system in:

H. T. Moges, T. Manos, Ch. Skokos (2022). Anomalous diffusion in single and coupled standard maps with extensive chaotic phase spaces. Physica D: Nonlinear Phenomena, 431, 133120. https://doi.org/10.1016/j.physd.2021.133120

Get Started

1. Install Dependencies

git clone https://github.com/htmoges/ChaosNetBench
cd ChaosNetBench
pip install -r requirements.txt
pip install -e .

2. Download the Dataset

Download the public benchmark dataset from Hugging Face:

pip install huggingface_hub
python -c "from huggingface_hub import hf_hub_download; \
    hf_hub_download(repo_id='htmoges/chaosnetbench-cml', \
    filename='data/chaosnetbench_cml.h5', repo_type='dataset', \
    local_dir='data/')"

For a quick local smoke test without downloading the full dataset, a mini subset is already included at data/chaosnetbench_cml_mini.h5.

Dataset schema, metadata, and Croissant records are documented in data/README.md.

3. Run a Quick Experiment

python scripts/train.py \
    --model graph_wavenet \
    --K 2.0 --rho 0.20 --N 8 \
    --seed 42

Or use the included mini dataset for a quick end-to-end test (no download needed):

python scripts/train.py \
    --model dlinear \
    --K 0.5 --rho 0.10 --N 8 \
    --seed 42 \
    --dataset data/chaosnetbench_cml_mini.h5

Included In This Release

• Core benchmark package: chaosnetbench/ (systems, dataset loading, metrics, models, training)
• Benchmark entry-point scripts: scripts/train.py, scripts/sweep.py, scripts/analyze_results.py
• Mini dataset for local smoke testing: data/chaosnetbench_cml_mini.h5
• Aggregate benchmark results: results/chaosnetbench_cml_results.csv
• Benchmark configuration: configs/benchmark.yaml

The full dataset (27.3 GB HDF5) is hosted on Hugging Face. Benchmark protocol details are in the paper and data/README.md.

Citation

If you use this code or dataset, please cite the associated paper.

@misc{moges2026chaosnetbench,
  title  = {ChaosNetBench: Benchmarking Spatio-Temporal Graph Neural Networks on Chaotic Lattice Dynamics},
  author = {Henok Tenaw Moges and Charalampos Skokos and Deshendran Moodley},
  year   = {2026},
    doi    = {10.48550/arXiv.2605.09676},
    url    = {https://arxiv.org/abs/2605.09676}
}

License

MIT License. See LICENSE.