I appreciate what they're doing. It's pretty cool looking. But
In this image it’s easy to see that the main centers of work are in cities, including San Francisco, Oakland, San Jose, and Sacramento, and that they are highly connected.
This image, too:
But where should planners draw the edges of a megaregion encompassing this activity?
Perhaps more importantly, why? You don't think the fact that Reno is across state lines has something to do with all the San Francisco traffic there? Also:
The results of the algorithmic analysis took some cleanup and iterations—such as eliminating superlong commutes between places like New York City and Los Angeles and excluding nodes with only very weak connections—to produce a coherent map of plausible megaregions.
Right. Like "cleaning up" the connections between Los Angeles and Las Vegas. I've literally driven there for meals. We're surrounded by Vegas billboards down here. Vegas exists for Los Angeles. Yet it's its own region because it's bigger than Reno, I guess.
A map shows all the commutes of 50 miles or less in the greater San Francisco Bay Area.
Right - a different one shows Phoenix as stretching north of the Grand Canyon.
Data visualization is a useful skill but I'm not convinced that their approach reveals any insight. There's clearly a lot of shaping going on. From the PLOS paper:
A heuristic approach is, put simply, one which involves exploration through trial and error to produce results which are useful, but not necessarily optimal. Simon  refers to this as the concept of “heuristic search” in which sub-optimal but effective “satisficing” solutions are arrived at. We take this visual approach with the ACS commuting and workplace data by working iteratively towards representations which identify core labor market areas based on networks of flows. In addition to providing a visual ‘sense check’ on the underlying data, this also provides a useful reference point for the algorithmic approach described later.
They took ZCTA data and fudged it until it looked cool. 'K. It looks cool. But if part of your method is "making it look cool" are we really learning anything from the process?