If Grid.face_lon does not exist, _populate_face_centroids(), actually _construct_face_centroids() in it, takes extremely long for large datasets. For instance, the benchmark/profiling below is for a ~4GB SCREAM dataset, around 5 mins:

@rajeeja FYI: I'm already working on this and have gotten optimized results, which will be good for "cartesian" parts of the face center calculations, but you may want to look into the Welzl parts as well, i.e. _populate_face_centerpoints().