You then use statistical analysis to identify the mathematical relationship between skull shape (as defined by the skull landmark coordinates) and facial shape (as defined by the facial landmark coordinates) across the database.

Once you have this statistical relationship, you can apply it to an unknown skull: given the skull landmark coordinates of the unknown individual, you use the statistical model to predict the facial landmark coordinates — and thus the facial shape — that would be most consistent with that skull, given the patterns observed in your reference database.

AFA3D extends this basic approach with additional sophistication, including the use of dense surface modeling rather than discrete landmark points, and the incorporation of additional anatomical constraints.

But the fundamental logic is the same: use statistical learning from a database of known skull-face relationships to predict the face from the skull.