LankaFlood AI

Research · Stage 4 of 8 done · No model trained yet

Can AI learn how floods evolve?

LankaFlood AI aims to teach a tiny world model how water moves, run it in your browser, and compare what it imagines with a physics simulation and with the real 2025 flood on the Kelani River in Sri Lanka.

Explore the real flood map Learn the maths

Three versions of one flood

The finished comparison puts these side by side. The world model doesn't exist yet.

Available now

Satellite mapping

What UNESCO and Copernicus mapped during the event. Open the map.

Available now

Simplified simulator

What simplified water equations produce. Not validated hydrology. Try a scenario.

Not built · arrives in stage 6

World model

What the trained model imagines, computed on your device.

The model at the centre

A world model learns the rule for what happens next. The experiments will run it both with and without weather inputs.

  1. sₜStateWater depth on each grid cell at time t.
  2. zTerrainElevation, plus editable obstacles and drainage paths. Fixed during a run.
  3. aₜWeather (optional)Rainfall, cloud and upstream inflow, if this version of the model uses them.
  4. ŝₜ₊₁PredictionWater depth one step later. It feeds back in as the next input.
Rollout: each prediction becomes the next state, so the model imagines 5, 10 or 20 steps ahead and its errors compound. That is what this project will measure.

What this tries to add

AI flood modelling is not new, and this is not the first such work for Sri Lanka. These are the working questions, narrowed by a first literature screen.

H1

How small can it get?

Measure how multi-step accuracy falls as the model shrinks to browser size, as a curve of size against rollout horizon.

H2

Which warning can a tiny model afford?

Larger flood models can already warn when they leave their training range, but the usual method multiplies download size. Test which cheap signal best predicts rollout failure on edited scenarios.

H3

Can it reach the next mapped state?

Train only on simulated floods, start from the observed 30 November state, and score on the observed 3 December state, only where both images saw the ground. It must beat a no-change baseline. Little rain fell between the two images, so this mainly tests draining. One case study, not proof.

H4

What is weather worth to a tiny model?

Train the same model with no weather, with rainfall, and with rainfall and cloud, then measure what each input adds per kilobyte. Not yet checked against the literature.

Roadmap

  1. 1
    Foundation

    Repository, static deployment and portfolio link.

    done
  2. 2
    Historical GIS data

    Real UNESCO and Copernicus flood layers on an aligned map.

    done
  3. 3
    Historical rainfall and terrain

    Half-hourly IMERG rainfall and Copernicus terrain added to the explorer.

    done
  4. 4
    Reference simulator

    A simplified 32 by 32 water-flow simulator, not validated hydrology, and synthetic training data.

    done
  5. 5
    World-model training

    Small PyTorch models trained with and without weather, adding cloud and inflow data.

    planned
  6. 6
    Browser inference

    The trained model exported to ONNX and run on the visitor's device.

    planned
  7. 7
    Research evaluation

    Validation against the real event, limitations and the article.

    planned
  8. 8
    Landslide extension

    A separate landslide head that reuses the terrain and weather inputs.

    planned