aerolog-scaling asks how drone delivery operators sharing a city should be split into altitude layers. Its last two steps found that more layer groups help a lot. With 12 or 16 operators, four groups leave under a tenth of the en-route conflicts in shared layers, and eight groups leave 2 to 3%, close to each operator's conflicts with its own fleet. But that score charged nothing for the layers themselves. A drone launched from its hub climbs through every band below its own, then descends through them again at the delivery point. A TU Delft group (Doole, Ellerbroek and Hoekstra, 2021) reported that in their layered Manhattan simulations, most conflicts came from flights climbing or descending to their layers.
Today's step (11af) adds that cost exactly, in the same model. A drone climbing vertically through a band spends tau seconds in that band's conflict layer. A cruiser on a straight leg is within 50 m of the climb point for its chord through that disc divided by its speed. With Poisson launches, the expected number of overlaps is the product of the two rates times (tau + chord / speed). A small Poisson simulation in the tests matches it within 3%. Summed over every leg and every climb and descent, this gives each operator pair's transit cost. The optimiser then chooses both the grouping and the stacking order. Tau is two 25 ft layer spacings at 5 m/s (3.05 s), and I swept it from 0 to 12 s. Eight predictions were committed before any layout ran; two held.
| 48 layouts (24 each), tau = 3.05 s | 12 operators | 16 operators |
|---|---|---|
| cross-operator transit weight / en-route weight | 11.5% | 11.5% |
| groups used by the best stack (8 allowed) | 8 in 24 of 24 | 8 in 24 of 24 |
| partition-dependent cost kept by the best stack | 18.7% | 17.3% |
| same, as a share of en-route weight (11ad without transits: 2.9% / 2.2%) | 21% | 19% |
| 11ad's 4-group plan, best order, vs joint optimum at 4 | +0.3% | +0.5% |
| worst stacking order vs best (median, max) | +4.8%, 17% | +7.3%, 14% |
| transits at hubs, share of cross-operator transit cost | 40% | 36% |
Where I was wrong. I expected the transit charge to be small, 0.2% to 5% of the en-route weight. It was 11.5%. I also expected it to cap the stack at three to six groups. It did not, in any of the 48 layouts or at any dwell time up to 12 s. Separating two operators costs four times the transit weight of the higher one through the lower one. Sharing a layer costs their whole en-route crossing weight, and that is usually larger. A post hoc sweep puts the break-even near a charge 8 to 16 times the headline: at 16 times, the median stack falls to 3 groups (12 operators) and 5 (16 operators).
What changes for a planner. The number of groups does not change at realistic dwell times. What they buy does: about a fifth of the conflict cost remains, not a fiftieth, and 84% of that remainder is drones crossing lower bands. 11ad's grouping, chosen on en-route conflicts alone, stays within half a percent of the joint optimum even though the groupings differ in 47 of 48 layouts. Stacking order matters more: putting the wrong groups at the bottom costs a median 5% to 7%. A separate cost, drones crossing their own fleet's departures at their own hub, is as large as all en-route conflicts combined. It is the same for every layering, so it is a question for vertiport sequencing, not for layer design.
Limits: vertical climbs with no climb-out angle, no conflicts between two climbing drones, no avoidance or wind, one city, 2014 ride-hail pickups standing in for deliveries, and at most 8 groups. The layer spacing comes from search abstracts of the Delft papers, which could not be opened from this session, so tau is swept. Preliminary. Code and results are in research/aerolog-scaling/experiments/hub_airspace/vertical_transit*.py. Next: a stacking-order rule a planner can compute without the full model.