Updraft | Devpost

Drones Traversing
ParaView model of South Kensington!
Stats

Inspiration

The de-facto approach to making actionable predictions in 2026 is to set up compute, find a dataset, and train an ML model. But this is boring, isn't it? Then we started thinking about something clearly more fun; Computational Fluid Dynamics (CFDs) 👴! With the rise of powerful hardware power in the age of AI, we believed you don't need to chase the next swanky algorithm to create valuable signal, so we put Kensington in a Wind Tunnel (sorry Imperial) to aid drone routing in dense urban areas.

What it does

Updraft models the travel of a drone through a real city's buildings, using their structure and effect on winds to optimise travel time and energy efficiency on long routes.

By comparing different strategies for route planning, we demonstrate that higher environmental awareness leads to more optimal routes. e.g. A drone slips into a tailwind between buildings to save on power.

How we built it

Set up our own high performance cluster!
Compiled OpenFoam and its dependencies, optimising it for our compute nodes.
Preprocessed metre-level accurate 3D models of London, extracting Kensington.
Ran dozens of CFD simulations, producing large datasets which we plugged into our app.
Used the data to multiple 3D pathfinding strategies to find best drone routes between two given points.
Developed visualisations of the simulated wind data on our model.
Created our beautiful design 🧑‍🎨!

Challenges we ran into

OpenFoam requires a custom scientific DSL, which we had to rapidly pick up (we don't do a physical science!)
Compiling OpenFoam required 200+ dependencies, which was challenging to reason with and parallelise.
Processing the high resolution wind data for a 3D search algorithm, which was a new problem to us.