Research

Neural dynamic programming for humanitarian logistics

In the first hours of a disaster, what responders learn about demand, road damage and stock depends on where they send scarce survey resources such as drones, and planning under this decision-dependent uncertainty is intractable at realistic scale. This project formulates relief logistics as a multi-stage, distributionally robust stochastic programme, solves it with stochastic dual dynamic programming (SDDP), and uses Neural SDDP to learn value-function approximations across related disaster scenarios and warm-start the solver. The methods will be tested on relief networks based on the 2010 Haiti earthquake and a Japanese earthquake case.

Human behaviour and human-drone interaction in disaster response

Evacuation plans rarely account for how people behave under stress, and relief models seldom capture how new information changes their decisions. This project uses virtual-reality experiments and surveys to study how people respond to sudden evacuation orders, how training affects their choices and why they deviate from safe routes. It then extends endogenous stochastic optimisation to human-drone interaction, modelling drones as multi-agent systems that locate and guide evacuees under changing conditions. The work includes the co-supervision of two PhD students and a Disaster Preparedness Workshop at Sabancı University.

Cold-chain optimisation of food supply chains in Cameroon

Perishable produce from smallholder farms loses quality quickly, and fragmented transport adds cost for farmers and traders. This project, a collaboration with AIMS Cameroon, redesigns the tomato supply chain from the Western Highlands (Foumbot and Mbouda) around a Farmer Cluster Center in Bafoussam, where farmers pool their produce before refrigerated transport to markets in Douala. A linear programming model allocates flows given transport costs, distances, supply and demand, with a quality degradation model that accounts for time, initial quality and temperature. Compared with the existing chain, the proposed design improved product quality and lowered costs.

• Food supply chains optimization considering cold-chains and food security (2024, AIMS Cameroon MSc essay; Chilonga)

Simulation of civil aviation networks and air traffic control

Weather and other disruptions force air traffic managers to rebalance demand and capacity across a whole network. This project built a modular, agent-based simulation of a civil aviation network and its control, combining machine-learning wind and convective weather prediction, 4D trajectory planning and a trajectory prediction model. Around this core it evaluated dynamic airspace sectorisation and developed an air traffic flow management model that assigns delays, reroutes, speed changes and cancellations, using UK airspace (the London and Scottish flight information regions) as the test case.

• A scalable reinforcement learning-based approach to dynamic airspace sectorization(2026, Transportation Research Part C: Emerging Technologies)
• Evaluation of Air Traffic Network Resilience: A UK Case Study(2024, Aerospace)
• Spatiotemporal Thunderstorm Forecasting for Pre-Tactical Air Traffic Operation: A Deep Learning Approach(2024, AIAA AVIATION FORUM AND ASCEND 2024)
• Temporally Correlated Deep Learning-Based Horizontal Wind-Speed Prediction(2024, Sensors)
• Enhancing air traffic operational efficiency by reducing network scale(2024, Aerospace Traffic and Safety)

Vertiport network design for urban air mobility

Where to place vertiports, how to size electric air taxis and how much infrastructure each vertiport needs are interlinked decisions that are usually taken separately. This project combines them in a three-stage model with a feedback loop, using queuing theory to size take-off, landing and charging pads. Applied to a London case study, it shows that ignoring vehicle sizing and vertiport queuing can produce networks that cannot operate, and that short waiting times are key to controlling costs. A companion study linked vertiport placement to demand, combining a binary logit mode-choice model, with taxi travel times from PTV Visum, and a hub location model for airport services connecting London and nearby cities with Heathrow and Gatwick.

• An integrated vertiport placement model considering vehicle sizing and queuing: A case study in London(2023, Journal of Air Transport Management)
• Demand modelling and optimal vertiport placement for airport-purposed eVTOL services (2023, TRB Annual Meeting; Serhal, Escribano Macias and Angeloudis)

Rapid evacuation of urban transport networks

Large-scale evacuations are often held back by congestion, imperfect coordination and damaged transport networks. This project developed a hybrid simulation-optimisation method that plans evacuations through demand staging and signal phasing, evaluating policies with a dynamic traffic assignment model that captures congestion, queuing and vehicle spillback, and searching for the best strategies with derivative-free optimisation. Combining departure time scheduling with signal phasing improved evacuation efficiency over a worst-case benchmark. A follow-up study models and optimises how drones can support evacuation management.

• Optimal design of Rapid evacuation strategies in constrained urban transport networks(2020, Transportmetrica A: Transport Science)
• Modelling and optimisation of drone-based support for evacuation management (2026, TRB Annual Meeting; Samuels et al.)

Target Level of Safety for unmanned aircraft systems

Before drones can fly beyond visual line of sight alongside manned aircraft, regulators need a target level of safety against which operations can be assessed. This study for the UK Civil Aviation Authority reviewed existing methods, classified UAS failure modes and sources of trajectory uncertainty, and proposed an agent-based digital twin of drone and manned air traffic. The approach assesses safety in segregated airspace, Transponder Mandatory Zones and non-segregated airspace.

• Agent-based Modeling Approach for Operational Risk Assessment of Large-scale Low-altitude UAV Traffic(2025, 2025 Integrated Communications, Navigation and Surveillance Conference (ICNS))
• Accuracy evaluation of a new generic Trajectory Prediction model for Unmanned Aerial Vehicles(2021, Aerospace Science and Technology)

Drone-assisted humanitarian relief

Drones can reach disaster areas quickly, but their range, battery management and fit with the wider relief supply chain limit how they are used. This project developed models that plan drone delivery of medical and relief supplies from field hubs, accounting for flight energy, battery management and fair distribution, and that coordinate drone surveys of road damage with ground relief vehicles when network conditions are uncertain. The models were tested on case studies based on the 1999 Chi-Chi earthquake in Taiwan and the 2010 Haiti earthquake.

• Optimal hub selection for rapid medical deliveries using unmanned aerial vehicles(2020, Transportation Research Part C: Emerging Technologies)
• Endogenous stochastic optimisation for relief distribution assisted with unmanned aerial vehicles(2020, OR Spectrum)
• Design of Hospital Delivery Networks Using Unmanned Aerial Vehicles(2020, Transportation Research Record: Journal of the Transportation Research Board)
• Integrated Trajectory-Location-Routing for Rapid Humanitarian Deliveries using Unmanned Aerial Vehicles(2018, 2018 Aviation Technology, Integration, and Operations Conference)