Does dark energy change character at a specific cosmic moment?

Standard neural network reconstructions of dark energy can produce physically nonsensical results because they chase data without constraints. Cosmo-PINN bakes the actual cosmological equations directly into the training loss, forcing solutions to obey physics at every step. Trained on DESI DR2 baryon acoustic oscillations, supernovae, and cosmic chronometers, it finds dark energy's equation-of-state crosses the so-called phantom divide — a boundary with implications for the universe's fate — around redshift 0.3, consistent with the standard parameterization but now with a physics-guaranteed reconstruction.