Machine learning reconstructs cosmic velocities with physics-aware shortcuts

Measuring the kinetic Sunyaev-Zel'dovich effect—which maps large-scale baryon distribution—requires accurately reconstructing galaxy velocities from spectroscopic data. Velocityformer, an equivariant graph transformer, deliberately breaks translation/rotation symmetries to match the observer's perspective, where the line of sight sets a preferred direction. The result: 35% better correlation with true velocities than physics-based methods, trained on just 4 simulations, and it generalizes across different surveys, cosmologies, and galaxy samples without retraining.