Orateur
Description
Machine learning techniques can be used to build near-optimal summary statistics for extracting cosmological information from the large-scale structure of the Universe. Their training, however, requires large numbers of simulated samples, making them expensive to construct. We aim to build a neural summary statistic trained on fast but approximate low-resolution simulations that remains informative when applied to higher-resolution ones, reducing the training-set cost. We train an ensemble of convolutional neural networks with different architectures to compress the matter overdensity field of low-resolution particle-mesh simulations into a few coefficients. The outputs of a subset of networks are concatenated to build a summary statistic, and we compare three selection strategies for the smallest informative sub-ensemble. We apply this neural summary to derive cosmological constraints on Ωm and σ8 from high-resolution particle-mesh simulations, using a Gaussian process emulator of the neural summary within a Gaussian likelihood, and compare it with a reference power spectrum analysis performed with the same procedure. We also consider dark matter halo catalogues from independent high-resolution N-body simulations. The neural summary extracts information beyond the binned power spectrum content when applied to higher-resolution data. Using multiple networks is essential, with a single network providing worse constraints than the power spectrum in several cases. Among the tested methods, a greedy search on the posterior volume provides the tightest constraints with the smallest subset of networks. The neural summary outperforms the power spectrum when applied to dark matter haloes, for different number densities and redshifts, despite being trained on low-resolution matter fields at z = 0. Using an ensemble of networks rather than a single one is therefore key, and offers a route to cheaper neural summaries.