Aetherscan

Breakthrough Listen's deep-learning SETI pipeline: a two-stage architecture where a Beta-VAE encoder compresses each observation of a 6-observation cadence (3 ON / 3 OFF, ABACAD) into an 8-dimensional latent, and a Random Forest classifies the cadence's concatenated latents as a technosignature candidate or not.

This repository carries the released model weights at stable filenames, versioned via git tags: training tags match the pipeline run's save tag (e.g. train_20260101_120000), and release tags (vX.Y.Z) mark blessed weights.

Training tag: train_20260729_152426

Files

File	Description
vae_encoder.keras	Beta-VAE encoder (Keras) — the inference feature extractor
vae_decoder.keras	Beta-VAE decoder (Keras) — for reconstruction/traversal analysis
random_forest.joblib	Random Forest cadence classifier (joblib)
config.json	Full resolved training configuration for this run

Training configuration

Parameter	Value
Training rounds	10
Epochs per round	100
Beta-VAE samples per round	499200
Random Forest samples	99840
Curriculum schedule	exponential
SNR base	10
Initial SNR range	40
Final SNR range	10
Latent dimensions	8
Beta (KL weight)	1.5
Alpha (clustering weight)	10.0
RF estimators	1000

The complete configuration is in config.json.

Evaluation (validation split)

Metric	Value
ROC AUC	1.0000
Average precision	1.0000
Classification threshold	0.99
Validation samples	19968

Library versions

Library	Version
python	3.12.3
tensorflow	2.17.0
numpy	1.26.4
scikit-learn	1.5.2
huggingface_hub	1.21.0

Usage

Pin this training tag with --hf-revision to download exactly these weights (a bare no-artifact inference download resolves to the latest vX.Y.Z release tag instead, never a training tag):

python -m aetherscan.main inference --hf-revision train_20260729_152426 --inference-files <catalog.csv>

Links & citation

Source code, documentation, and issue tracker: https://github.com/zachtheyek/Aetherscan. If you use Aetherscan in your research, please cite it via the repository's CITATION.cff.