Technology

Flood modelling that learned from physics

A 2D flood model solves the shallow-water equations over millions of cells, thousands of times a storm, which is why a full ARR study usually takes weeks. Flash trains a neural network on those solutions, so the answer for a new site and storm comes back in about a second, then checks it against the physics.

1. Physics first

Everything starts with a GPU 2D shallow-water solver: a well-balanced finite-volume scheme that keeps still water still, wets and dries cells cleanly, applies Manning friction implicitly, and takes rainfall losses and impervious surfaces cell by cell. It runs dozens of storms at once on one graphics card. It is the reference the AI learns from and is checked against, and it runs the design storms of every certified study.

2. Learning from physics

The neural network is a U-Net, the architecture used to read satellite and medical images, conditioned on the storm. It reads the terrain (elevation, slopes, depressions, where water collects, roughness and hard surfaces) and the rain (every five minutes for up to nine hours, with losses) and draws the peak depth and speed at every cell in one pass.

  • Trained on tens of thousands of full physics simulations of ARR design storms
  • Real Queensland LiDAR and synthetic valleys, hills and urban blocks, at 1, 2, 2.5 and 5 m cells
  • Pre-burst rain, initial and continuing losses, roads, roofs and paved areas
  • Large sites are split into overlapping tiles and blended seamlessly
  • Runs on an ordinary computer's processor: no GPU, no cloud queue

3. Checking it

Every version of the network is scored against the physics solver on storms it has never seen, and on real Queensland sites kept well away from any terrain it trained on. We measure how closely the flooded area and the depths match, and we keep refining: larger networks, more simulations, more real terrain.

We don't ask you to take the AI on trust. Studies run their design storms through the physics solver, and an engineer reviews every result before it reaches you.

4. Where the AI is used, and where it isn't

Site previews and site forecasts are the network alone: fast, and for a first look only. Flood studies use it to explore options quickly, run the design storms through the physics solver, and are reviewed by an engineer; certified studies are signed by a Registered Professional Engineer of Queensland.

The full toolkit

AI-powered flood modelling
A deep neural network trained on tens of thousands of full 2D flood simulations over real Queensland LiDAR and synthetic terrain. It predicts peak depth and speed for a storm in about a second on an ordinary computer: no GPU, no queue, no waiting weeks.
GPU 2D physics solver
The full shallow-water equations on graphics processors, dozens of storms at once: a well-balanced finite-volume scheme with proper wetting and drying. It trains and checks the AI, and runs the design storms of every certified study.
ARR 2019 ensemble engine
Every AEP, duration and temporal pattern the guidelines ask for, with pre-burst rain, losses and areal reduction; median of the patterns and the critical duration at every spot.
Live Australian data
Queensland Government LiDAR at 1 m, BoM design rainfall and the ARR Data Hub fetched for your exact site, with land use and soils for losses.
Hazard, impact and GIS outputs
ARR hazard classes H1 to H6, pre- and post-development flood impact, flood outlines and GeoTIFF grids ready for your GIS or CAD.
Drainage structures
Culverts, open channels, floodways, basins and pit and pipe networks to Austroads AGRD05, with blockage and climate change factors.
Benchmarked against the industry
Everything is modelled in-house on our own engines, and tested against the published 2D benchmark tests and established industry models.
Engineers in the loop
Every certified study is reviewed and signed by a Registered Professional Engineer of Queensland.

Under the hood: Tomioka is the platform, Urokodaki the physics solver, Tanjiro the neural network.

Try it free on your site See the products