Data-Driven Discovery of a Simple Phantom-Crossing Dark Energy Parametrization

Abstract

We develop a data-driven reconstruction programme for the dark-energy equation of state within VCDM, a minimally modified gravity framework in which both background and linear perturbations can be consistently evolved across the phantom divide. Using CMB, BAO, and type-Ia supernova data, we first perform a Bayesian spline reconstruction of $w(a)$, finding a preference for smooth, monotonic phantom-crossing trajectories. Bayesian evidence disfavors increasingly complex spline models, indicating that current observations exhibit a statistical preference for low-complexity dark-energy dynamics. Motivated by this result, we apply Exhaustive Symbolic Regression, an interpretable machine-learning technique that systematically searches over analytic expressions of fixed complexity, identifying the remarkably simple one-parameter form $w(a)={w_0}/{\sqrt a}$, which reproduces the reconstructed behaviour and fits the data at a level comparable to standard two-parameter parametrizations such as CPL. The model naturally crosses the phantom divide for $w_0<0$, suppresses early dark energy, and predicts a transient accelerating and phantom phase without a future big-rip singularity. As a one-parameter model, it is highly predictive, being a genuinely dynamical deformation of the cosmological constant rather than containing it as a limit. Bayesian model comparison yields mild-to-moderate support for this parametrization relative to standard two-parameter alternatives, and stronger evidence relative to $Λ$CDM. Our results suggest that current observations favour surprisingly simple dark-energy dynamics and illustrate how Bayesian reconstruction and symbolic regression can be combined into a principled model-discovery framework for cosmology.