Foundation Models for Exoplanet Characterization

This project aims to build a multimodal foundation model for exoplanet characterization by addressing the challenge of heterogeneous astronomical data and instrument-specific biases. It combines transformer-based models for images (protoplanetary disks) and sequential data (transit light curves), trained using knowledge distillation, masked autoencoding, and contrastive learning on both real and simulated datasets. The deliverables include pretrained multimodal models, a curated data pipeline, evaluation on real-world tasks, and a well-documented open-source codebase with reproducible workflows.