Ly$\alpha$NNA: A Deep Learning Field-level Inference Machine for the Lyman-$\alpha$ Forest

TL;DR

This paper introduces a deep learning framework using a residual CNN to infer IGM thermal parameters from the Lyman-$\alpha$ forest at the field level, outperforming traditional summary statistic methods in constraining power.

Contribution

The authors develop a novel deep learning approach that directly models the Lyman-$\alpha$ forest spectra, significantly improving parameter inference over classical summary statistics.

Findings

Deep learning yields ~11-fold improvement over power spectrum alone.

The framework accurately estimates parameter covariances and posterior distributions.

Perfect data analysis shows substantial information gain from field-level inference.

Abstract

The inference of astrophysical and cosmological properties from the Lyman-$\alpha$ forest conventionally relies on summary statistics of the transmission field that carry useful but limited information. We present a deep learning framework for inference from the Lyman-$\alpha$ forest at field-level. This framework consists of a 1D residual convolutional neural network (ResNet) that extracts spectral features and performs regression on thermal parameters of the IGM that characterize the power-law temperature-density relation. We train this supervised machinery using a large set of mock absorption spectra from Nyx hydrodynamic simulations at $z=2.2$ with a range of thermal parameter combinations (labels). We employ Bayesian optimization to find an optimal set of hyperparameters for our network, and then employ a committee of 20 neural networks for increased statistical robustness of the network inference. In addition to the parameter point predictions, our machine also provides a self-consistent estimate of their covariance matrix with which we construct a pipeline for inferring the posterior distribution of the parameters. We compare the results of our framework with the traditional summary (PDF and power spectrum of transmission) based approach in terms of the area of the 68% credibility regions as our figure of merit (FoM). In our study of the information content of perfect (noise- and systematics-free) Ly$\alpha$ forest spectral data-sets, we find a significant tightening of the posterior constraints -- factors of 10.92 and 3.30 in FoM over power spectrum only and jointly with PDF, respectively -- that is the consequence of recovering the relevant parts of information that are not carried by the classical summary statistics.