Symbolic Emulators Like CMBolic Could Reshape Cosmic Structure Analysis for Next-Gen Surveys
Preprint demonstrates fast, interpretable symbolic emulators for CMB lensing that match neural network accuracy while enabling rapid Bayesian inference, with potential to uncover patterns missed by traditional methods.
The CMBolic preprint (arXiv:2606.07745, not yet peer-reviewed) introduces analytic symbolic emulators for the CMB lensing potential power spectrum in an extended ΛCDM model incorporating massive neutrinos and CPL dark energy. Unlike neural network emulators such as those in CosmoPower (arXiv:2202.07589), which require trained weights and offer limited interpretability, CMBolic delivers closed-form expressions achieving 0.27-0.32% mean absolute fractional error on validation spectra up to ℓ=5500. Methodology relies on symbolic regression over parameter spaces, validated against CLASS Boltzmann code outputs rather than full N-body simulations. This approach slashes inference runtime from weeks to minutes when paired with ACT DR6 and Planck lensing likelihoods, yet the study tests only a narrow extended model subspace and omits full non-linear structure formation effects that future surveys like CMB-S4 will probe. Related work in symbolic regression for cosmology (e.g., arXiv:2305.13342) shows these methods can reveal hidden parameter degeneracies neural nets obscure, potentially exposing evolving dark energy signals hidden in current numerical pipelines. Limitations include reliance on linear-theory approximations and the need for broader validation across dynamical dark energy variants; if scaled, symbolic emulators may become essential for real-time Stage-4 data processing where neural opacity risks systematic biases.
Helix: Symbolic emulators could replace black-box neural nets in cosmology pipelines, offering interpretable shortcuts that accelerate Stage-4 survey analysis while exposing hidden dark energy and neutrino signatures.
Sources (3)
- [1]Primary Source(https://arxiv.org/abs/2606.07745)
- [2]Related Source(https://arxiv.org/abs/2202.07589)
- [3]Related Source(https://arxiv.org/abs/2305.13342)