Classifying modified gravity and dark energy theories with Bayesian neural networks: massive neutrinos, baryonic feedback, and the theoretical error

ABSTRACT We study the capacity of Bayesian neural networks (BNNs) to detect new physics in the dark matter power spectrum. As in previous studies, the Bayesian Cosmological Network (bacon) classifies spectra into one of five classes: $\Lambda$CDM, $f(R)$, $w$CDM, Dvali–Gabadaze–Porrati (DGP) gravity, and a ‘random’ class, with this work extending it to include the effects of massive neutrinos and baryonic feedback. We further develop the treatment of theoretical errors in bacon-ii, investigating several approaches and identifying the one that best allows the trained network to generalize to other power spectrum modelling prescriptions. In particular, we compare power spectra data produced by euclidemulator2, hmcode, and halofit, all supplemented with the halo model reaction to model beyond-$\Lambda$CDM physics. We investigate BNN classifiers trained on these sets of spectra, adding in Stage-IV survey noise and various theoretical error models. Using our optimal theoretical error model, our fiducial classifier achieves a total classification accuracy of $\sim$95 per cent when it is trained on euclidemulator2-based spectra with modification parameters drawn from a Gaussian distribution centred around $\Lambda$CDM ($f(R)$: $\sigma _{fR0} = 10^{-5.5}$, DGP: $\sigma _{r\mathrm{c}} = 0.173$, $w$CDM: $\sigma _{w0} = 0.097$, $\sigma _{wa}=0.32$). This strengthens the promise of this method to glean the maximal amount of unbiased gravitational and cosmological information from forthcoming Stage-IV galaxy surveys.