Kaluza-Klein reduction of 11-dimensional supergravity on $G_2$ manifolds yields a 4-dimensional effective field theory (EFT) with $N=1$ supersymmetry. $G_2$ manifolds are therefore the analog of Calabi-Yau (CY) threefolds in heterotic string theory. Since 2017 machine-learning techniques have been applied extensively to study CY manifolds but until 2024 no similar work had been carried out on $G_2$ manifolds. We first show how topological properties of these manifolds can be learnt using simple neural networks. We then discuss how one may try to learn Ricci-flat $G_2$ metrics with machine-learning.