Networks of neurons on graphs in astrophysics: a review finds that only one in six articles apply them to real observations

Christian Kragh Jespersen defends the GNN as the most natural architecture for cosmology, while noting that they serve mainly on simulations.

A galaxy in its environment, a halo of dark matter in its history, a photon impact in a detector: according to Christian Kragh Jespersen, physics advances by studying objects in their relationships, and graphs are the ideal structure to describe them. In a 43-page review, a guest chapter of Springer's collective book on machine learning in astrophysics and cosmology, he argues that networks of neurons on graphs (GNNs) often adapt better to physical data than other architectures: they respect the geometry of the problem, integrate known symmetries and remain more interpretable.

The author reconstructs these networks on the basis of the principles of deep geometric learning, showing that convolutive networks and transforming them are special cases. His overview covers cosmology, galaxy formation, dynamics and instrumentation, and provides a finding: apart from instrumentation, only one out of six articles apply these methods to real data rather than simulations. The journal also proposes good practices, including when not to use a graph.