jiazhe868/sensoformer-data

Sensoformer datasets

Preprocessed, model-ready HDF5 datasets for Sensoformer: amortized inference of earthquake moment tensors and magnitudes from variable-geometry station sets.

File	Events	Size	Role
socal_mxyz_data_rtz_lp2_ampr_ps_wlola.hdf5	2,435	0.26 GB	real Southern California catalog (M ≥ 3.0) with analyst mechanisms — fine-tuning and evaluation
syn_mt_data_realgeom_realvn_10w_ps_wcoda_wlola.hdf5	~100k	11.4 GB	PSDR synthetics on real station geometries — pre-training
syn_mt_data_noaug.hdf5	~50k	5.4 GB	clean synthetics, no randomization — the ablation baseline

Format

Per event, one HDF5 group:

/<event_id>/
    waveforms      (S, 12, 101) float32   # P/S windows Z,R,T + their amplitude spectra
    features       (S, 20)      float32   # distance, azimuth, station lon/lat, depth,
                                          # P/S max amplitudes, spectral maxima, log P/S ratios
    station_names  (S,)         vlen str
    attrs: magnitude, Mxx, Myy, Mxy, Mxz, Myz, depth [, evlo, evla, strike, dip, rake, grade, date]

S (station count) varies per event. Windows are 5 s before to 5 s after each arrival at dt = 0.1 s; horizontals are rotated to radial/transverse; waveform normalization is per event so inter-station relative amplitudes (which carry the radiation pattern) are preserved. Moment tensors are normalized deviatoric components with Mzz = −(Mxx + Myy); magnitude is scaled to [−1, 1] over [2, 8] at load time. Full specification: DATA_FORMAT.md.

Usage

pip install git+https://github.com/jiazhe868/sensoformer.git
python scripts/download_assets.py --datasets socal-real
python scripts/predict.py --input socal-real --out-dir results/

import h5py
with h5py.File("socal_mxyz_data_rtz_lp2_ampr_ps_wlola.hdf5") as f:
    g = f[list(f)[0]]
    print(g["waveforms"].shape, g["features"].shape, dict(g.attrs))

Provenance

• Real data: waveforms from the Southern California Earthquake Data Center (SCEDC), with analyst focal mechanisms from the Yang–Hauksson–Shearer catalog. Downloaded and preprocessed with the scripts in scripts/data_acquisition/ and scripts/preprocessing/; filtering 0.2–2.0 Hz.
• Synthetics: frequency-wavenumber synthesis on real station geometries, with Physics-Structured Domain Randomization — velocity model sampled per event from a CRUST1.0-derived library, stochastic travel-time and amplitude perturbation, injected real ambient noise, synthetic scattering coda, and station dropout. Filtering 0.1–2.0 Hz. The clean dataset omits all randomization.

The preprocessing in the repo reproduces these files byte-for-byte from the raw archives (verified by tests/test_preprocessing_agreement.py).

Please cite the SCEDC and the source-mechanism catalog in addition to the Sensoformer paper when using the real data.

Citation

@article{jia2026sensoformer,
  title   = {Sensoformer: Robust Sim-to-Real Inference on Variable-Geometry
             Sensor Sets via Physics-Structured Randomization},
  author  = {Jia, Zhe and Zhang, Xiaotian and Li, Junpeng},
  year    = {2026}
}