UniverseTBD/platonic-universe

Empirically tests whether foundation models trained on different astronomical data modalities (optical imaging, infrared, spectroscopy) converge toward shared representations, using mutual k-nearest neighbor alignment metrics across Vision Transformers, DINOv2, ConvNeXtv2, IJEPA, and astronomy-specific models. Combines multi-modal astronomical datasets (HSC, JWST, DESI, Legacy Survey) with pretrained vision encoders to measure representational collapse at scale, providing evidence that larger models exhibit stronger cross-modal alignment. Includes CLI tooling and Python APIs for running end-to-end embedding generation and comparative analysis workflows.