Soojung Yang analyzes whether foundation models trained on molecules, materials, proteins, and language converge toward a shared representation of matter. She identifies both convergence and failure modes, then shows how representational alignment can transfer knowledge between scientific models and improve learning across tasks.