đź’ˇInspiration
Wildfires hit harder every year, and the tools meant to predict them just aren’t keeping up. Static maps, outdated indexes, slow updates, it all leaves people reacting after the damage is already done. We wanted something that could look ahead, not behind. So we built F.I.R.O: a system that warns you about danger before it turns into disaster.
🌍What it does
F.I.R.O builds a digital twin of any region using real environmental data: wind, humidity, temperature, vegetation, drought levels, elevation, and more.
With it, F.I.R.O can Simulate fire behavior:
Run what-if scenarios (wind shifts, humidity drops, new ignition points) and watch the model update in real time.
If a fire is active, F.I.R.O shows how it will spread and helps guide decisions like:
- where firefighters should move,
- where aircraft should drop water,
- which areas should evacuate.
⛰️Why Digital Twin
We chose a digital twin because resolution fundamentally determines whether a model reflects reality or averages it away. This difference is enormous operationally: fires don’t spread uniformly; they jump ridges, slow down in moist gullies, or accelerate up slopes. Escape routes and safety zones depend on those small-scale features. Wind flow near terrain is highly variable: 10 m grids can capture those local eddies and slope winds that 10 km grids completely miss.
So, unlike coarse 10 km models that smooth entire landscapes into single pixels, a 10–30 m digital twin offers the ability to simulate fire behavior at the scale people and vehicles actually operate.
🛠️How we built it
We are using Unity3D and Cesium to create a high-resolution digital twin with real-world terrain, elevation, and satellite data. Cesium enables meter-level geographic accuracy, allowing simulations to reflect real environmental conditions instead of coarse grid approximations.
A regression-based ML model was trained on historical wildfire and environmental data (wind, humidity, temperature, vegetation, slope). After normalization and spatial alignment, the model was integrated directly into Unity for real-time inference.
This setup allows the system to dynamically update fire risk and spread based on changing inputs, enabling interactive what-if simulations and accurate visualization of fire behavior at operational scale.
⚠️Challenges we ran into
- Getting consistent, accurate historical data.
- Normalizing wildly different datasets so the model wouldn’t freak out.
- Making predictions fast enough to actually feel realtime.
- Designing an interface that stays simple even though the system behind it is pretty complex.
🏆Accomplishments that we're proud of
- Built a lightweight, optimized 3D digital twin in Unity that maintains real-time performance (30+ FPS) even on mobile devices. The system uses efficient terrain streaming, LOD management, and optimized rendering pipelines to ensure smooth interaction without sacrificing spatial accuracy or simulation fidelity.
- Built a working ML model that produces reliable risk predictions.
đź“šWhat we learned
- ML is only as good as the data you feed it... and the data for wildfires is pretty chaotic.
- Predictive tools need constant updating, something that isn't yet fully possible with current technologies.
- Cross checking model outputs with real historical cases helped us fine tune everything way more than expected.
🚀What's next for F.I.R.O
AR support for firefighters: We’re exploring AR headsets that let firefighting crews see the live fire spread prediction directly in their field of view, wind direction, danger zones, and potential fronts layered over the real world.
Crowdsourcing App: A companion mobile app will let citizens submit photos, and hazards in a fire prone area. These inputs feed back into FIRO to make the prediction model faster and more accurate.
Bringing in satellite imagery and real-time weather feeds to make predictions even sharper.
Collaborating with organizations to test FIRO in real-world field conditions.
Log in or sign up for Devpost to join the conversation.