Ze-Yu Peng, Qing-Yu Lan, Yun-Song Piao

Oscillatory features in the primordial power spectrum, potential signatures of new physics in the early universe, are usually searched for using fixed templates. In this work, we perform a template-free search for primordial features using symbolic regression. We find that both Planck and the combined Planck+ACT+SPT-3G datasets independently select an inverse-$k$ oscillation, $\cos(B/k)$ with $B\simeq4\,\mathrm{Mpc}^{-1}$, as the leading low-complexity feature. Comparing this inverse-$k$ template with standard linear and logarithmic oscillating templates, we find that it fits the data best, showing a weak preference for a non-zero amplitude. Our results show that symbolic regression as a powerful machine learning technique can provide an interpretable, model-independent approach to cosmological discovery.