Machine Learning and more specifically Deep Learning has successfully erupted
into a vast number of engineering fields in the recent years, specially leaping
traditional simulation approaches by leveraging data usage. Even though the
potential is huge, the delicate selection of an adequate Machine Learning
architecture for a specific problem determines the success of its
implementation. This is essential for non-Euclidean datasets, like the ones
found in social networks, molecule structures, manifolds, and others. In those
datasets, the distance between two points does not correspond to the Euclidean
distance, but to the path along the edges (either weighted or unweighted). This
is the case of Computational Fluid Dynamic (CFD) meshes. In all these fields,
the fitting of Graph Neural Networks (GNNs) for this type of datasets have made
them gain popularity in the recent times. Specially as aerodynamic predictors
they have had a remarkable dominance during the last few years, as not only
there is a strong academic research trend toward these architectures, but many
“AI-consulting engineering companies” offer them as the surrogate model of
choice. In this survey, a brief introduction to GNNs is presented. More
importantly, and different from other GNN surveys, this review paper focuses on
their current application as aerodynamic coefficients and flow field predictors
(academic and industrial), with emphasis on their specific architecture.
Nineteen publications have been selected for this review, focusing, but not
exclusively, on external aerodynamics.