Oracle-2-Image_Backbone_ZTF

ZTF Image Backbone for BTS

ORACLE: Real-time, hierarchical photometric classifier for ZTF (BTS) / LSST (ELAsTiCC). Part of Oracle Collection.

Code: github.com/dev-ved30/Oracle

Docs: https://dev-ved30.github.io/Oracle/

Papers: ORACLE-1 2501.01496 | ORACLE-2 2607.00228

Model Details

• Architecture: e.g. GRU_MD_Improved: Bi-GRU (hidden 128x2 layers) + attention pooling -> LayerNorm/GELU/Dropout(0.2) + static MLP -> residual head -> 16-d latent -> hierarchical logits (n_nodes).
• Inputs: [change per variant]
  • Lite: ts (B,T,5)=[flux, fluxerr, wavelength, mjd-first_det, photflag] + length
  • Standard: Lite + static (B,30 BTS / 18 ELAsTiCC) host/context metadata
  • Omni: Standard + postage_stamp (B,3,224,224) [science, reference, difference]
• Outputs: Conditional probs per sibling group -> class probs per node via taxonomy.get_class_probabilities(). Use model.predict(table) / model.score(table) -> {level: label}.
• Loss: Weighted Hierarchical Cross-Entropy (WHXE).
• Training data: BTS = ZTF Bright Transient Survey real alerts; ELAsTiCC = ELAsTiCC2 LSST simulations. Truncated by days_since_trigger.

Intended Use

Real-time triage on alert streams from 1 observation onward. Hierarchical outputs allow high-confidence decisions at Level 1 (Transient vs Variable / Persistent vs Transient) even when leaf is uncertain. Demonstrated live on ZTF stream.

Out-of-scope: Not for spectra, not for anomaly classes outside taxonomy (see Limitations).