Orthogonally, I wonder if these arrangements are generalizable to give lower bounds for N = n^2 + 1
Also increasing the grid may be helpful to understand which are atoms, if some ideal atoms get split to fix in the grid, and if the bars in the center are real or an artifact. (Is there a 45° rotated square in the center?). So this also may help to understand and generalize the result. But the run time grows very fast for big grids, so I'm less optimistic about this possible improvement.
But where is the picture for this lower bound packing?