Airlift Planner

Airlift Planner estimates, for any earthquake in the USGS catalog, which nearby airfields can still take relief flights, how much cargo the hardest-hit population needs each day, and how much of that an airlift could move. It uses only public data: USGS ShakeMap and PAGER, US Census population files, and the OurAirports airport table.

It is a student project and a planning model, not an operational tool.

Dashboard: https://sachabidermann.github.io/airlift-planner/

USGS HayWired scenario, a simulated M7.0 on the Hayward Fault. Red rings: airfields the model flags as likely knocked out. Blue square: the recommended gateway, San Francisco International. Shading: USGS ShakeMap intensity. Base map © OpenStreetMap contributors.

Quick start

Requires uv. The first run downloads about 30 MB of public data into data/; the full check brings it to about 100 MB.

git clone https://github.com/sachabidermann/airlift-planner
cd airlift-planner
uv sync
uv run main.py --quake gllegacyhaywiredm7p05_se     # any USGS event or scenario id
uv run main.py --list                                # M5.5 and above in the past week
uv run main.py --latest                              # plan for the newest one

Abridged output for the HayWired scenario (... marks lines left out; terms are defined below):

M 7.0  Haywired M7.05 Scenario  (USGS scenario, simulated)
shaking: USGS ShakeMap v33, 2017-01-11
...
people at MMI VIII+     5.4M   (airbridge priority)
cargo needed           6,103 t/day   at 1.13 kg per person per day

nearest airfield     KLVK Livermore Municipal Airport            15 km   MMI X    usable 19%
likely knocked out:
    KLVK Livermore Municipal Airport            15 km   MMI X    usable 19%
    KHWD Hayward Executive Airport              17 km   MMI IX   usable 37%
    KSJC Mineta San Jose International Airpo    26 km   MMI IX   usable 24%
    KOAK Oakland San Francisco Bay Airport      28 km   MMI IX   usable 27%
    ...

GATEWAY   KSFO San Francisco International Airport        38 km   MMI VII  usable 92%
          747-8F x 6 spots, 11,870 ft   inflow 3,160 t/day
...
AIRLIFT CAPACITY INTO ZONE    3,160 t/day   (ceiling for the assumed parking spots; t = metric tons)
people it could supply         2.8M
share of need                   51%   the rest must come by road, sea or local supply

Check it

scripts/check.sh --offline    # unit tests and the browser-model test; no network, a few seconds
scripts/check.sh              # also reruns the backtests, the verification and the dashboard build

VERIFY.md says what each step proves and what output to expect. The browser-model test needs Node.js. GitHub runs the offline checks on every push.

Terms

• MMI: Modified Mercalli Intensity, a I to XII scale of how hard the ground shook at one place; ShakeMap reports values up to X. VI is strong shaking with light damage; VIII is severe with heavy damage.
• ShakeMap: the USGS map of estimated MMI around an earthquake, published as a grid and revised as data arrives.
• PAGER: the USGS estimate of how many people were exposed to each MMI level.
• Airbridge: a sustained shuttle of cargo flights into a disaster area.
• Gateway: the large airport, outside the worst damage, where heavy cargo jets land.
• Forward strip: any usable airfield inside the damage zone, from a small strip to an international airport, served by C-130s from the gateway.
• Parking spot: room for one cargo aircraft to unload. How many an airfield can work at once (the Air Force term is MOG, maximum on ground) limits throughput more than runway length does.
• Sortie: one round trip by one aircraft between the gateway and a forward strip.
• Usability: the model's probability that an airfield can take relief flights in the first three days. Below 50% it is flagged as likely knocked out.
• Damage center: the average position of the people shaken above MMI VI, weighted by how far above. All distances are measured from it.
• Zone: airfields within 150 km of the damage center, or shaken at MMI 6.5 or more.
• t/day: metric tons (1,000 kg) per day.

How it works

Formulas and the source of every constant are in METHOD.md.

1. Usability. Read the ShakeMap intensity at every airport within 1,000 km of the damage center and map it to a usability probability. The airfield nearest the damage is often among the hardest shaken, so nearest is a bad rule. This step replaces it.
2. Need. Take the people at MMI VIII and above, from PAGER, or from a Census reconstruction where PAGER was not run. Multiply by a daily ration of food, medical supplies and a share of a shelter kit: 1.13 kg per person per day.
3. Airbridge. Heavy jets fly into a gateway; C-130s shuttle cargo on to forward strips. Throughput uses the formula in Air Force Pamphlet 10-1403: parking spots × payload × operating hours ÷ ground time × 0.85, then multiplied by the airfield's usability. The inputs used here are more generous than the pamphlet's own examples; METHOD.md gives the comparison. Every candidate gateway is worked out in full. The one that moves the most into the zone wins, except that candidates within 10% of the best count as tied and the nearest of them wins. Gateways in the affected country are preferred.
4. Result. Airlift capacity into the zone in t/day, how many people that could supply, and the share of need.

module	job
airlift/usgs.py	Earthquake records, ShakeMap grids, PAGER exposure; real events and scenarios
airlift/population.py	Census exposure where PAGER was not run
airlift/airports.py	Airports and runways
airlift/survivability.py	Intensity to usability
airlift/demand.py	Exposure to metric tons per day
airlift/aircraft.py	Aircraft payloads, runway needs, ground times
airlift/airbridge.py	Damage center, gateway choice, shuttle allocation, capacity
airlift/report.py, main.py	Terminal report and command line
backtests/	Event definitions with sources, the backtest runner, the verification script, and their generated reports
dashboard/	Exports dashboard data and tests the browser port of the model
docs/	The static dashboard served by GitHub Pages
tests/	Unit tests

Backtests

Nine real earthquakes, compared with what happened. Generated detail, including the model's values at each airport that mattered and sources for every historical statement, is in backtests/RESULTS.md.

event	model flags as likely knocked out	what happened to airports	model gateway	relief gateway used	match
Loma Prieta 1989, M6.9	none	Oakland lost 3,000 ft of runway to liquefaction	San Jose	none; mostly road	n/a
Anchorage 2018, M7.1	none	none; reopened the same day	Anchorage	none	n/a
Ridgecrest 2019, M7.1	none	China Lake not mission capable	Mojave	none	n/a
Puerto Rico 2020, M6.4	none	none	San Juan	none	n/a
Haiti 2010, M7.0	none	tower unusable, field saturated	Port-au-Prince	Port-au-Prince	yes
Nepal 2015, M7.8	none	runway later damaged by heavy jets	Kathmandu	Kathmandu	yes
Turkey 2023, M7.8	Kahramanmaras (48%), Hatay (34%)	Hatay closed six days	Gaziantep	Adana and Incirlik	no
Morocco 2023, M6.8	none	none	Marrakech	Marrakech	yes
Myanmar 2025, M7.7	Mandalay, Nay Pyi Taw, 3 other fields	Mandalay and Nay Pyi Taw closed to commercial flights for a week; relief flights from day 2 and day 4	Heho	Yangon	no

Gateway matches in 3 of the 5 events where a relief gateway was used. The other four earthquakes had no sustained relief airlift, so they only test the flags.

This is weak evidence. The usability curve was set by hand while looking at the five international earthquakes, so for those this is a consistency check, not an out-of-sample test. The US events were added later without changing the curve. In the three matches the airport used was the obvious one: the capital's international airport in Haiti and Nepal, and the city airport beside the damage in Morocco. Nine earthquakes is a small sample either way.

Flags. Three airports closed after the shaking and the model flagged all three: Hatay for six days, and Mandalay and Nay Pyi Taw for a week to commercial flights, though military relief flights reached Nay Pyi Taw two days after the earthquake and Mandalay after four. Judged strictly on relief flights in the first three days, Hatay and Mandalay are hits, Nay Pyi Taw is a miss, and Kahramanmaras at 48% is a false alarm: it stayed open to relief flights. Three other Myanmar fields were flagged and I could not find what happened to them.

Misses. Oakland in 1989 lost 3,000 ft of runway to liquefaction at MMI 6.7, where the curve gives 92%; intensity does not capture soft ground. China Lake in 2019 was declared not mission capable, and the model gave it 64%.

In Turkey and Myanmar the model picks an intact airport close to the damage. Responders used a larger airport farther away. The model cannot explain that choice because it does not see customs, fuel or cargo handling. Gaziantep did take relief flights in 2023.

Scenarios

Scenario earthquakes from the USGS scenario catalog, run through the planner unchanged.

scenario	people at MMI VIII+	gateway	flagged as likely knocked out	airlift capacity	share of need
HayWired M7.0, San Francisco Bay Area	5.4M	San Francisco	9, including Oakland, San Jose, Hayward and Livermore	3,160 t/day	51%
ShakeOut M7.8, Los Angeles	8.5M	Long Beach	16, including Ontario, San Bernardino, Chino and Fullerton	3,168 t/day	32%
Cascadia M9.0, Pacific Northwest	136k	Portland	none; tsunami is not modeled	3,070 t/day	exceeds need
Seattle Fault M7.5	2.4M	Paine Field	4: Sea-Tac, Boeing Field, Renton, Bremerton	1,252 t/day	46%
New Madrid M7.7, Memphis	306k	Memphis	10 airfields in Arkansas, Missouri and Tennessee	3,138 t/day	exceeds need
Puerto Rico Trench M8.5	2.1M	San Juan	none	2,110 t/day	88%

Seattle and New Madrid use USGS PAGER exposure. The other four use the Census reconstruction, which counts US residents only.

In all six scenarios the gateway is within 50 km of the damage center, so all of its cargo counts as delivered and capacity equals what the gateway can receive. The shuttle fleet only decides how much of that cargo is flown forward. The share of need therefore rests on the assumed parking spots.

Verification

uv run backtests/verify_shaking.py writes backtests/VERIFICATION.md:

1. Our grid is read at 4,155 cities from the USGS PAGER city lists for nine events and compared with the intensity PAGER assigned. For eight events the mean difference is 0.16 or less in absolute value. Anchorage is +0.44. Its ShakeMap was revised 489 days after PAGER ran, which may explain part of it, but five other events were also revised later and agree within 0.16, so the cause is open. The largest single-city difference is 1.53. This checks that the grid is read correctly. It says nothing about whether ShakeMap itself is right.
2. The parser for the older grid.xml format agrees with the current format at 500 random points: mean difference 0.004, largest 0.39.
3. The fallback distance formula, used only until USGS publishes a ShakeMap, is off by 0.70 intensity units on average across 1,519 airports, and by 0.49 where MMI is V or above.
4. The Census reconstruction, run on events that also have PAGER, gives 0.87 of PAGER's count at MMI VIII and above for the Seattle scenario and 1.05 for New Madrid. It does worse on smaller events. At MMI VII and above the ratios are 0.47 for Anchorage, where the Census point for a very large municipality sits 33 km from the city, and 0.64 for Ridgecrest. At MMI VIII and above it put 29,161 people where PAGER has 1,691 for Puerto Rico 2020. Small counts are unreliable.

The browser port of the model is checked against the Python planner on 135 cases (node dashboard/test_model.js).

Limitations

• Capacity is a ceiling for an assumed number of parking spots (6 at a large gateway, 3 at a medium one), not a forecast. Every limit the model leaves out can only lower it. The spot counts are guesses and can be wrong in either direction: a cargo hub such as Memphis parks far more than six freighters.
• Forward strips are picked on runway length and width. Pavement strength and ownership are not checked, so private and turf strips appear in the lists.
• The usability curve is a judgment call, not a fitted model. FEMA's Hazus has published fragility curves that should replace it.
• Distances are straight lines.
• The model ignores liquefaction, tsunami, fuel supply, customs, ground handling, weather, airspace limits and road conditions.
• The airport table and Census population are current, so replays of older events can include airports that did not exist and use 2023 population.
• OurAirports changes nightly and USGS revises ShakeMaps, so results are tied to their inputs. RESULTS.md records the input hashes and ShakeMap versions.

Live events

ShakeMap is revised repeatedly, sometimes years later. The tool caches the version it first downloads and prints the version and date. During a response, re-run with --refresh to fetch the latest. Before USGS publishes a ShakeMap, the tool falls back to a distance formula and says so in the output.

Dashboard

docs/ runs the same model in the browser. Pick an event and change the assumptions (shuttle fleet, operating hours, forward radius, minimum usability, domestic only). The plan, map, charts and tables update.

uv run dashboard/build.py                      # refresh docs/data.json and dashboard/expected.json
node dashboard/test_model.js                   # the browser model must reproduce the Python planner
python3 -m http.server 8766 --directory docs   # http://localhost:8766

?theme=light or ?theme=dark forces a theme. The URL hash selects an event, for example #wa22sfz01_se.

Data and credits

Earthquakes, ShakeMap and PAGER: U.S. Geological Survey (public domain). Population: U.S. Census Bureau, vintage 2023 population estimates and 2023 Gazetteer files (public domain). Airports: OurAirports (public domain). Base map: © OpenStreetMap contributors, ODbL. Map library: Leaflet (BSD 2-Clause). This project is not affiliated with or endorsed by any of them.

Related work

USGS ShakeCast sends facility-level shaking alerts from ShakeMap to operators such as Caltrans. FEMA's Hazus has airport fragility curves and reads ShakeMaps. A 2021 CISA and Washington State study rated the runways of 20 airports against the USGS Cascadia M9.0 scenario. Air Force Pamphlet 10-1403 gives the airfield throughput formula used here. I have not found a public tool that goes from a ShakeMap to an airfield ranking, a relief need and an airbridge plan.

Palantir's disaster work with Direct Relief, Team Rubicon and the World Food Programme integrates an organization's own data into one operating picture. A model like this could run on top of that.

Next steps

Replace the hand-set curve with the Hazus fragility curves, then add liquefaction susceptibility for runways on fill. After that: tsunami inundation for coastal strips, road travel time in place of straight-line credit, a helicopter tier for areas without a strip, and fuel and handling limits on gateway choice.

Sacha Bidermann, 2026.