Sensoformer — pretrained weights

Set-attention inference of earthquake moment tensor and moment magnitude from a variable-size set of seismic station recordings.

• Code, docs and CLI: https://github.com/jiazhe868/sensoformer
• Datasets: https://huggingface.co/datasets/jiazhe868/sensoformer-data
• Architecture: dual-tower 1D-ResNet station encoder → 3-layer transformer encoder over stations (128-d, 4 heads, no positional encoding) → additive attention pooling → magnitude and moment-tensor heads. 2,063,015 parameters.

Files

File	Stage	Use for
sensoformer_v3_finetuned.pth	PSDR synthetic pre-training + real fine-tuning	inference on real data
sensoformer_v3_psdr_pretrained.pth	PSDR synthetic pre-training only	fine-tuning on a new catalog or region

Each file is a self-describing checkpoint: {state_dict, arch, sensoformer_version, stage, metrics}.

Usage

pip install git+https://github.com/jiazhe868/sensoformer.git

from sensoformer import load_pretrained
model = load_pretrained("sensoformer-v3-finetuned", device="cuda")  # downloads from here
predictions, attention = model(waveforms, features, mask)
# predictions: (B, 6) = [scaled Mw, Mxx, Myy, Mxy, Mxz, Myz];  Mw = (y+1)/2*6 + 2

Inputs must follow the documented HDF5 / tensor layout: per station a (12, 101) waveform block (P and S windows, Z/R/T, plus their amplitude spectra) and 20 scalar features (distance, azimuth, station lon/lat, depth, amplitude summaries). See DATA_FORMAT.md.

End-to-end CLI:

python scripts/predict.py --input my_events.hdf5 --out-dir results/ --catalog

Evaluation

244 held-out real Southern California events (M ≥ 3.0; seed=42, 80/10/10 split):

Metric	Value
Median Kagan angle	19.7°
Mean Kagan angle	23.9°
Magnitude MAE	0.100
Fraction with Kagan < 30°	0.77

The median equals the analyst catalog's own median 1σ nodal-plane uncertainty (19.5°), i.e. the model is at the label noise floor. Baselines under the identical protocol: MPNN 24.3°, DeepSets 29.1°, DeepONet 28.5°, no pre-training 25.2° (median Kagan). Full tables: RESULTS.md.

Intended use and limitations

Intended for research on amortized seismic source inversion and as a starting point for fine-tuning on other catalogs/networks.

• Trained on M ≥ 3.0 Southern California events. Below that range magnitudes floor near Mw ≈ 2.7, giving a roughly constant positive bias (+0.28 at M2.5–3.0, +0.71 at M2.0–2.5); mechanisms degrade only gradually (22.5° and 28.3° median). Remove the constant offset, or fine-tune with small events, before using magnitudes out of range.
• Error is geometry-limited: median Kagan ranges from ~17° for well-surrounded events (azimuthal gap < 45°) to ~33° beyond 180°. Stratify by azimuthal gap or station count rather than quoting a single number.
• Region/network transfer is untested beyond Southern California. The architecture is geometry-agnostic, but re-fine-tuning and re-calibration are recommended.
• The model does not locate events; hypocenter and origin time must come from a catalog or locator.
• Reported conformal intervals (90%: ±0.22 magnitude, 44.8° Kagan ball) are calibrated for this catalog population — recalibrate elsewhere.

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}
}