Real water distribution network data is private. Utilities won't share pressure readings, flow rates, or demand patterns — it's security-sensitive infrastructure. But machine learning for leak detection, demand forecasting, and network optimization requires extensive observations. The field is data-starved because the data exists behind institutional walls.
DiTEC-WDN (arXiv:2503.17167) builds 228 million synthetic hydraulic states across 36 water distribution networks: 36,000 unique scenarios simulated over 24-hour and 1-year periods. An automated pipeline optimizes pressure, flow rate, and demand parameters, then validates each scenario against physical rules to ensure hydraulic realism.
Note (2026-11-14; read against the paper's full text): the essay says the pipeline's validation step checks that every synthetic state obeys conservation of mass, energy equations and pressure constraints. In the paper the validation is a set of rules that judge whether each simulated outcome is valid, introduced to catch anomalies such as negative pressure or time inconsistency; the states themselves come from a hydraulic simulation (EPANET input files), and the paper does not describe the validation step as a check of mass or energy conservation. It also says the data cover normal conditions only, with anomalies such as leakage, fire-fighting and pipe break excluded, and that the dataset is built on 36 publicly available networks (10 simulated over 24 hours, 26 over a year, 1,000 scenarios per network), which the essay's '36,000 scenarios over 24-hour and 1-year periods' compresses.
The dataset supports graph-level, node-level, and link-level regression plus time-series forecasting. Each network snapshot is a graph — nodes are junctions, edges are pipes, attributes are pressures and flows. This graph structure is the natural representation: water networks ARE graphs, and the physics respects the topology.
The structural insight: privacy restrictions that block data sharing create a market for synthetic data that must be physically realistic without being physically real. The pipeline's validation step — checking that every synthetic state obeys conservation of mass, energy equations, and pressure constraints — is doing the work that real data does for free by existing. The synthetic data must earn its realism through explicit physics, while real data gets it automatically. This makes the synthetic pipeline, paradoxically, more transparent than the real data it replaces: every assumption is encoded in the generation rules, nothing is hidden.