Robust neural network-enhanced estimation of local primordial non-Gaussianity

When applied to the nonlinear matter distribution of universe, neural networks have been shown be very statistically sensitive probes cosmological parameters, such as linear perturbation amplitude ${\ensuremath{\sigma}}_{8}$. However, when used a ``black box,'' are not robust baryonic uncertainty. We propose architecture for constraining primordial non-Gaussianity ${f}_{NL}$, by training network locally estimate ${\ensuremath{\sigma}}_{8}$, and correlating these local estimates with large-scale density field. apply our method $N$-body simulations, show that $\ensuremath{\sigma}({f}_{NL})$ is 3.5 times better than constraint obtained from standard halo-based approach. has same robustness property halo bias: physics can change normalization estimated but cannot whether ${f}_{NL}$ detected.