Cold hydrogen in the Milky Way: a lightweight neural network detects self-absorption at 21 cm, at a rate of 15,000 spectra per second

Trained on synthetic spectra, this convolutional neural network identifies the signatures of cold atomic gas without modelling the underlying emission, with 96.5 per cent accuracy, at a rate compatible with the future SKA.

Cold atomic hydrogen is the intermediate step between the hot, diffuse gas of galaxies and the molecular clouds where stars are born. Its best tracer, when no background source is available, is the self-absorption visible in the emission line at 21 cm. But extracting it systematically remains difficult: existing methods are based on eye inspection or emission modelling, two approaches that are difficult to reproduce on a large scale.

A network with no hypotheses about the show

Researchers on arXiv propose a lightweight convolutional neural network that detects these signatures and deduces their speed through a posteriori processing, without hypothetizing the emission or additional data. Trained on simulated spectra, it achieves 96.5% accuracy, 96.0% precision and 97.0% recall on discarded synthetic data.

Applied to real-world observations of the Riegel-Crutcher cloud and giant molecular filaments of the galactic plane, it regains the distribution and velocities of already known structures, in agreement with previous analyses and carbon monoxide data. Importantly, it processes about 15,000 spectra per second on the GPU of a simple laptop, paving the way for real-time detection at the expected rates of the Square Kilometre Array.