Transient Mode

Transient Mode

Transient Mode uses YOLOv8 (You Only Look Once) to detect and localize transient events in astronomical images.

What It Detects

• Supernovae (stellar explosions)
• Novae (thermonuclear explosions on white dwarfs)
• Variable stars (brightness-changing stars)
• Other transient phenomena

Performance

v1.0 (Initial Model)

• Trained on TNS + ZTF transient images (~2,000 images)
• mAP50: 51.5%
• Precision: 42.5% | Recall: 61.2%

v1.1 (Fine-Tuned During Streaming)

• Trained on 2,742 images with 5,234 bounding box annotations
• mAP50: 99.5% (+48 points)
• Trained autonomously during the 3-day streaming discovery run
• Model auto-retained the best checkpoint from epoch 3

Training Pipeline

The transient pipeline runs in 3 automated phases:

Phase 1: Data Collection

• Downloads transient images from TNS (Transient Name Server)
• Fetches from ZTF (Zwicky Transient Facility)
• Creates YOLO-format bounding box annotations
• Validates and deduplicates images

Phase 2: YOLO Training

• Model: YOLOv8-nano (lightweight, fast inference)
• Epochs: up to 100
• Image size: 128x128
• Batch size: 16
• Augmentation: Mosaic, flip, erasing
• Best weights saved based on validation mAP

Phase 3: Integration

• Best model copied to inference directory
• Tests on existing anomaly candidates
• Reports confirmation rate

Data Sources

Source	Type	Coverage
TNS (Transient Name Server)	Confirmed transients	Global, curated
ZTF (Zwicky Transient Facility)	Survey alerts	Northern sky

Pre-trained Models

Two models ship with AstroLens (in models/ via Git LFS):

• yolo_transient_v1.pt -- Initial model (51.5% mAP50)
• yolo_transient_v1.1.pt -- Fine-tuned model (99.5% mAP50)