Time / Space — Airport reconstruction

Loading dataset

CSV format

Use a comma-separated file with a header. Each row is one nonstop flight observation in one direction.

Column	What to enter
`from_airport` Required	Departure airport code, e.g. `LHR`.
`to_airport` Required	Arrival airport code, e.g. `JFK`. Must differ from departure.
`minutes` Required	Positive duration in minutes, up to 2,880. Decimals are allowed.
`aircraft` Optional	Aircraft model, e.g. `Boeing 777-300ER`.
`aircraft_code` Optional	Equipment code, e.g. `77W`. Used with the aircraft model for Jets only.

Example rows:

from_airport,to_airport,minutes,aircraft_code
LHR,JFK,480,77W
JFK,LHR,420,77W

Repeated observations are averaged within each direction. If both directions exist, their means receive equal weight. One direction is enough.

Use 4–300 airports in one connected network. The default minimum is 4 connections per airport; lower it in Flight filters for a smaller network.

Jets only keeps recognized jet aircraft. Without aircraft data, All flights is selected. Geographic and other extra columns are ignored. Uploaded flights are assumed to be nonstop; the importer does not verify them.

Download complete example CSV

File rules

UTF-8, comma-separated, up to 32 MB and 100,000 observations. Include a header and the same number of fields on every row. Quote values containing commas; escape a quote by doubling it. Airport codes may contain 2–12 letters, numbers, underscores or hyphens. Codes are converted to uppercase.

AIRPORT INSPECTION

The number beside each airport is its total edges to distinct airports in the filtered network. Both directions count as one edge. Counts include connections hidden in this camera view. Targets use scheduled durations, including individually audited corrections. No uniform overhead is subtracted; corrected rows retain the original provider duration and evidence. Synthetic distances are unchanged. Expand a row for directional times, optimizer weight and sources. Triangle exclusions appear below retained constraints with their witnesses; excluded edges do not contribute to the total-edge counts.

Airport / total edges	Target h	Fitted h	Error

Earlier fit with the duration cutoff

With minimum connections disabled and the former ≥3-hour cutoff, Lagos (LOS) had 17 pairs and 53 filtered observations, using original flight durations. After 30,000 steps, random start 71 placed it at radius ratio 0.28 with 10.6% local distance error. Start 72 reached ratio 0.97 with 2.6% local error using exactly the same times. Re-fitting Lagos alone with its neighbors fixed could not repair the first result: the whole fit is in a worse local minimum.

In that previous fit, Brasília (BSB) and Surabaya (SUB) each had one retained neighbor. Their distance could fit almost perfectly while their locations remained undetermined. Changing random starts cannot resolve missing measurements.

Casablanca → Lagos and Dubai → Lagos were checked against their explicit nonstop schedule tables; retained aircraft and durations match. These checks do not independently verify actual flight times. The surface passes through every fitted airport, including inward dents. Radius ratios describe a fitted sphere for diagnosis only.

Deepest points in the current fit

Distance from the best-fit sphere center, as a fraction of its radius. This diagnostic does not constrain the shape. Points in dents still belong to the all-airport surface.

Download pocket investigation

About this experiment

Airports start at random 3D positions. The optimizer moves them to match flight-time distances. No latitude, longitude, sphere or plane is supplied.

The synthetic sphere and flat-plane experiments use exact distances to show that either shape can emerge from the same optimizer.

Real flights include winds, routing and airport overhead. We average both travel directions when available. Jets only is the default to reduce aircraft-speed differences.

Distance error measures the fit. Radial variation measures roundness; plane thickness measures flatness. These shape measurements describe the result; they never guide the optimizer.

Data & downloads

Airport observations CSV Collection & nonstop audit Excluded flight observations Triangle & short-flight trial Excluded constraint evidence

Synthetic datasets

Sphere distances Sphere ground truth Sphere benchmark results Plane distances Plane ground truth Plane benchmark results

Earlier investigations

Outlier corrections Shape investigation SJO & Canada fit SSA investigation Interior points

Flight schedule sources

Flight routes cross-checked with Virtual Radar Server standing data. Flags from FlagCDN. Select an airport to inspect individual schedules and correction evidence.