Understand · Stage 7

What is new here?

There are no model results yet. This page records what the project is trying to find out, so the claims can be checked against the outcome later.

Working questions

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.

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.

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.

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.

Status: narrowed by a first literature screen in November 2026. It was a short search, not a systematic review, so "not found" does not prove a gap. Any question found to be answered elsewhere will be replaced.

Prior work so far

From the first screen. Not every paper here has been read in full yet.

Predicting Urban Flood Risk for the Kelani River BasinSouth Eastern University of Sri Lanka, 2025
Water-level forecasting at Hanwella with models including XGBoost. Not a spatial world model.
CNN-Weighted Cellular Automaton for Urban FloodingInternational Journal of Disaster Risk Science, 2024
Very close to the architecture considered here, so it is a baseline to beat or borrow from.
Kelani River Basin Flood Impact AnalysisSri Lanka Survey Department, 2026
Interactive analysis of observed impacts. No learned dynamics.
Assessment of Surrogate Models for Flood InundationWater Research, 2024
Shows learned models struggle in unfamiliar conditions, which motivates H2.
Geometry-Informed Neural Operator surrogate of LISFLOOD-FP, ForlìGarofalo et al., arXiv, 2026
A long rollout tested on real events, using one 2.7 million parameter model on a GPU. It starts from spin-up, not a mapped flood.
Uncertainty-aware reduced-order flood modellingJacquier et al., arXiv, 2020
Ensembles already warn when a flood surrogate leaves its training range, so H2 must be about affordable warnings.
Neural Wind TunnelOpen-source project on GitHub
A surrogate of about 0.7 MB in the browser, for steady airflow. It has no time-stepping, and clamps sliders to its training range.

Known limits

One real transition is a test case, not a training set. The model must learn from simulation.
Copernicus depth is modelled and its total includes extrapolation, so only the observed layer comes directly from imagery.
Far more ground was not analysed on 3 December, so dates can only be compared fairly where both saw the ground.
River inflow and any dam operations over those three days are unknown so far.
Results using measured rain are hindcasts, and will look better than real forecasts.
Rainfall comes from near-real-time public imagery, not gauge-corrected data, and it is coarse: about 11 km cells.
Almost no rain fell between the two flood images, so the real test mainly measures draining.
Terrain includes trees and buildings, which can block or misdirect simulated water.
The simulator is simplified and unvalidated, with no infiltration, inertia or river channels. A model that matches it has learned the simulator, not the river.
A rollout that matches this one event would not show that the model generalises.

Ready for landslides

The same storm also caused landslides. UNESCO maps about 2,163 ha of them. That layer is downloaded and measured, and the design keeps terrain and weather inputs separate from each hazard. No landslide model exists. It needs its own study area and its own mathematics, as the guide explains.

Roadmap

1
Foundation
Repository, static deployment and portfolio link.
done
2
Historical GIS data
Real UNESCO and Copernicus flood layers on an aligned map.
done
3
Historical rainfall and terrain
Half-hourly IMERG rainfall and Copernicus terrain added to the explorer.
done
4
Reference simulator
A simplified 32 by 32 water-flow simulator, not validated hydrology, and synthetic training data.
done
5
World-model training
Small PyTorch models trained with and without weather, adding cloud and inflow data.
planned
6
Browser inference
The trained model exported to ONNX and run on the visitor's device.
planned
7
Research evaluation
Validation against the real event, limitations and the article.
planned
8
Landslide extension
A separate landslide head that reuses the terrain and weather inputs.
planned