AI Framework for Jordan's Water Network: NRW Proof-of-Concept

Mohammed Fasha and four coauthors published an arXiv paper on 14 June 2026 presenting an AI-driven framework that pairs physics-based hydraulic simulation with LLM-based agents to address non-revenue water in Jordan. The paper frames the problem by noting that 50% of water produced in Jordan is lost to leakage, theft and metering issues, and offers a proof-of-concept run on a 1,164-junction Amman district network.

How does the framework detect leaks and anomalies?

The framework combines real-time data streams, EPANET hydraulic modeling, a digital twin layer and LLM-based AI agents to detect anomalies, then issues control actions via function calling. The authors use flow-based anomaly detection that aligns with water distribution zone practice, retrieval-augmented generation (RAG) for policy interpretation, and function calling for network control; the implementation uses EPYT for hydraulic simulation and offline LLMs (llama3.1:8b via Ollama).

The system ingests SCADA and sensor feeds into the digital twin, runs automated hydraulic simulations to model expected flows, and compares measured flows to simulated baselines to flag deviations. On detection, the LLM agents generate health reports and suggested actions, and can trigger control routines through the function-calling interface.

What did the Amman proof-of-concept show?

The proof-of-concept validated technical feasibility on a 1,164-junction Amman district network and produced automated hydraulic simulation, flow-based anomaly detection, and AI-generated health reports with response times under 2 minutes and zero API costs. In a simulated burst scenario, a 30.1 L/s leak produced measurable flow redistribution across 15 pipes and led the system to flag a 15-junction cluster that localised the burst, confirming alignment with distribution zone monitoring practice.

The implementation relied on EPYT for offline simulation and ran LLMs locally via Ollama, which the authors note kept operational costs at zero API spend while delivering sub-2-minute response times for report generation and control decision output.

Why does this matter for Jordan and similar water-scarce regions?

Jordan faces severe water scarcity and extremely high non-revenue water losses; the paper cites that 50% of produced water is lost. Embedding physics-based modeling with AI agents offers continuous monitoring and adaptive decision-making rather than purely reactive responses, which the authors argue traditional approaches have failed to sustain. The proof-of-concept shows the approach can localise a burst quickly in a complex urban network and produce operational outputs that fit existing water distribution zone practices.

The use of offline LLMs and zero API costs speaks to feasibility in contexts with constrained budgets and intermittent connectivity, while the phased implementation model described in the paper accounts for limited automation and intermittent supply common in Jordan.

What to watch

Adoption steps and field trials beyond the EPYT offline simulation are the next concrete signals: a live deployment that connects the framework to operational SCADA control loops would test whether the sub-2-minute response and zero-API model hold under production constraints. The authors flag phased implementation to accommodate intermittent supply and limited automation; the timelines and results of such phases will determine whether the approach moves from technical feasibility to operational impact.

The paper supplies several concrete data points for evaluation: 50% NRW in Jordan, a 1,164-junction proof-of-concept network in Amman, sub-2-minute report response times, zero API costs, and a simulated 30.1 L/s leak that redistributed flow across 15 pipes and flagged a 15-junction cluster. These figures frame the immediate benchmarks any follow-up implementation must match or exceed.