Uncertainty-Aware Deep Learning for the Lyα Forest: CNN-Based Absorber Detection and Characterization

The Lyα forest is a powerful probe of the intergalactic medium and small-scale matter distribution, but deriving absorber properties traditionally requires computationally expensive Voigt-profile fitting. We present a convolutional neural network (CNN) that identifies and characterizes H i Lyα absorbers directly from quasar spectra. The model is trained on synthetic spectra generated from the IllustrisTNG simulation and fitted with the viper Voigt-profile fitting code to provide training labels. The network simultaneously predicts absorber presence, column density (NHI), Doppler parameter (bHI), and line centroid. On simulated spectra, the CNN reproduces the viper reference catalogue with an F1 score of ∼0.8, with mean absolute errors of ∼0.18 in log NHI and ∼0.10 in log bHI. It reproduces the H i column density distribution function (CDDF) and the bHI–NHI relation obtained from viper, recovering similar CDDF slopes and a lower-envelope relation with an RMS difference of only 0.36 km s−1. Applied to high-resolution UVES spectra, performance decreases to an F1 score of ∼0.5, with mean absolute errors of ∼0.35 in log NHI and ∼0.22 in log bHI. Latent-space analysis reveals a significant domain shift between the simulated and observational spectra, contributing to the reduced performance. Nevertheless, the CNN reproduces the observed CDDF and bHI–NHI distributions obtained from the viper reference catalogue, yielding CDDF slopes consistent with viper and a lower-envelope RMS difference of 3.20 km s−1 relative to viper. Monte Carlo dropout is implemented during inference to quantify predictive uncertainties. Together with its computational efficiency, the method provides a scalable and uncertainty-aware framework for Lyα forest analysis in upcoming spectroscopic surveys.