Deep dive · Simulation & operations research
After the earthquake, Port-au-Prince Airport became the lifeline for an entire country: a plane landing every 17.5 minutes, one runway, and never enough water, food, or medicine getting through. We rebuilt those six weeks as a discrete-event simulation to answer the question the responders faced in real time: where does the next forklift, parking space, or ground crew actually buy the most relief?
17.52 min
mean aircraft inter-arrival
30 x 45
replications x scenario days
4
capacity configurations tested
+67%
water delivered (6,000 to 10,000 tons)
0.0456
SSE of the exponential arrival fit
For a month and a half, one airport had to absorb the world's aid: military, government, and NGO aircraft with different priorities, hard-scheduled slots against medical flights that could not wait, and diversions whenever the field hit capacity. The Air Force and the international aid community needed to know which constraints actually bound the system, and what adding capacity would have bought.
My role was data collection: sourcing the landing logs and pressing the client for the data the model needed. From there, we computed inter-arrival times in Jupyter from every completed flight, dropped pre-event records, and treated gaps longer than four hours as off-duty periods rather than demand signal. The cleaned series averaged 17.52 minutes between landings, and an exponential distribution fit best (SSE 0.0456 across candidate distributions), consistent with memoryless arrivals.
The diagnostic plots held one story worth telling: the series was stationary until it suddenly was not. Our best explanation was the seaport reopening, which would divert traffic the airport had been carrying alone. If so, the anomaly in the data was the relief effort working.
A discrete-event simulation in Simio: aircraft as entities carrying type (military, government, NGO), priority, and cargo attributes; the runway, taxiway, offloading ramps, forklifts, parking, and ground crews as constrained resources; diversions triggered at capacity. Validation ran through iterative client reviews: inter-arrival and ground service-time distributions moved from exponential to uniform to match the client's observed operations, crew counts were rebalanced, and the model's delivered-resource totals were checked against the client's records, which confirmed the real shortfalls in water, food, and medical supplies.
We ran 30 replications of the 45-day scenario across four configurations. The interesting ones did not assume instant infrastructure: a new parking space every four days (two at most) and one forklift plus five ground crew added per day for the first six days, modeling what a real response could actually stand up mid-crisis. The higher arrival-acceptance rate, by contrast, switched on at once.
Every supply category rose 44 to 67 percent with no changes to the runway itself: the tested configurations added ground capacity, parking, forklifts, and crews, alongside higher arrival acceptance. For future humanitarian assistance and disaster relief operations, the recommendation is flexible resource allocation planned in advance, so surge capacity exists on day one instead of day twenty, with larger infrastructure like bigger runways left to future study.