Challenge
Engineers planned upgrades by gut feel and spreadsheets. A handful of manually tested scenarios for a network worth nine figures of investment. Every intervention changes how water flows downstream — upsize a pipe here, and flooding shifts to a junction three miles away. Without a way to search the full space of options, planning was slow, expensive, and conservative.
The test catchment alone contained 7,886 pipes, 8,040 junctions, 47 pumps, and 1,829 subcatchments. Under a 30-year design storm, the baseline network generated 52,000m³ of flooding. Engineers manually trialling a handful of scenarios couldn’t come close to searching a space this large.
The planning problem isn’t picking the right pipe. It’s searching thousands of combinations where every upgrade changes what happens downstream.
Network ModelHydraulic model + constraints
AI OptimisationGenetic algorithm explores upgrades
Interactive DashboardExplore trade-offs, pick a strategy
Approach
The optimiser doesn’t pick a single best answer — it maps the entire trade-off curve between cost and flooding. Each generation breeds 50 candidate upgrade combinations, runs every one through a full hydraulic simulation, and feeds the results back. Better solutions survive, weaker ones are replaced, and the search converges on the frontier of what’s possible.
5,425 eligible pipes reduced to 1,359 groups. Grouping pipes by network topology and engineering rules makes the search space tractable without losing hydraulic accuracy.
Before handing the problem to the genetic algorithm, a Design of Experiments stage seeds the search with 50 strategically chosen scenarios. Each sampling strategy explores a different region of the solution space — from extreme max-upgrade configurations to engineering-informed starting points.
The optimiser ran for 49 generations after the initial experiments — 2,500 total scenarios across a 3-machine workgroup over 54 hours. Each scenario creates a full network configuration, runs a hydraulic simulation under the 30-year design storm, and extracts flooding and cost results.
The tool handles dropped connections, non-convergent simulations, and storage cleanup automatically. It runs unattended over a weekend — no babysitting, no manual restarts.
Planners explore results through an interactive dashboard — filter the Pareto front by cost range or asset type, drill into individual solutions, push any scenario back into the hydraulic model for detailed review. A heatmap of the most commonly upgraded assets surfaces no-regret investments: upgrades that appear regardless of which trade-off point you choose.
Result
57% flooding reduction. One weekend of compute. £100M of infrastructure investment optimised.