Contrastive Learning: Aligning Chandra Event Data with Research Papers
Key Highlights
- •Collaborating with Center for Astrophysics (CfA) Harvard on cutting-edge contrastive learning research
- •Creating shared embedding spaces between Chandra X-ray Observatory event data and astronomical research papers
- •Developing novel encoder architecture that bridges observational data with scientific literature
- •Enabling direct retrieval and semantic search across both data and text modalities
Overview
Collaborative research project with the Center for Astrophysics at Harvard to create shared embedding spaces between Chandra X-ray Observatory event data and astronomical research papers using contrastive learning.
Challenge
Astronomical research involves two distinct modalities that traditionally exist in separate silos:
- Observational Data: Raw event files with time, energy, and position measurements
- Scientific Literature: Research papers describing and interpreting observations
The goal was to create a unified embedding space enabling cross-modal retrieval between data and text.
Approach
Developed contrastive learning framework with dual encoders:
- Event Encoder: Processes Chandra event lists using CNN/Transformer hybrid
- Text Encoder: Processes research papers using pre-trained language models (SciBERT)
- Contrastive Loss: NT-Xent loss to align embeddings in shared space
Used hard negative mining and temperature scaling for efficient training on paired observation-paper datasets.
Technologies
- PyTorch: Model development and training
- Python: Core implementation and data processing
- RunPod: GPU infrastructure
- Pinecone: Vector storage for similarity matching
Impact
Achieved >85% top-5 retrieval accuracy for paper-to-data matching, enabling intelligent data retrieval, literature-guided analysis, and automated annotation of astronomical observations. Embeddings cluster meaningfully by astronomical object type and generalize to unseen source types.