EPANET-Agentic: A multi-agent system for natural language-controlled simulations of water distribution networks
EPANET-Agentic: Мультиагентная система для управления моделированием сетей водоснабжения на естественном языке
2026-01-21
SCID: 54.1/qrntux5j
Discuss with AI
EPANET-Agentichuman-in-the-loop oversightmulti-agent systemnatural language-controlled simulationwater distribution networks
Figures from the paper
Abstract (AI)
Water distribution networks (WDNs), a critical part of urban infrastructure, normally require numerous model simulations for effective planning and management. However, traditional WDN modelling requires complex workflows and specialized expertise. EPANET is the most widely adopted modelling tool for WDN hydraulics and water quality simulations, yet its operational complexity restricts accessibility and slows timely decision-making. Recent advances in large language models (LLMs) have led to the development of agentic artificial intelligence systems that autonomously coordinate tasks and control complex engineering simulations through natural language prompts. Here we introduce EPANET-Agentic, a multi-agent system that integrates advanced workflow reasoning with the EPANET simulator and incorporates human-in-the-loop oversight for critical interventions. The new platform adopts an orchestrator-centred, tool-driven architecture that nests three specialised agents (TaskExecutor, CodeRunner, and DataAnalyzer) as function-call tools. This design enables autonomous task decomposition, precise tool invocation, and transparent workflow management. The abilities of EPANET-Agentic are evaluated on three benchmark networks (i.e., L-Town, C-Town, and Net3) across four categories of tasks: System Characteristics, System Dynamics, System Operation, and Scenario Simulation. The results demonstrate that EPANET-Agentic achieved a 100% success rate and tool invocation accuracy with no human interventions. Moreover, the multimodal DataAnalyzer agent provided valid interpretations of simulation results, while the nested tool design ensured robustness and the architecture exhibited strong scalability across diverse hydraulic analysis tasks. These findings confirm that EPANET-Agentic enables natural language-controlled WDN simulation and analysis with engineering-grade reliability, while still adhering to a human-in-the-loop approach required for safety-critical systems. With its modular architecture and strong adaptability, EPANET-Agentic marks a step change from conventional WDN modelling approaches, positioning itself as a next-generation platform for complex planning and management challenges.
Key Findings
1
Across L-Town, C-Town, and Net3 benchmark networks, EPANET-Agentic achieved 100% task success and tool-invocation accuracy without human interventions.
2
EPANET-Agentic integrates large-language-model workflow reasoning with EPANET simulations through an orchestrator-centered, tool-driven multi-agent architecture.
3
The architecture demonstrated robustness and scalability while retaining human-in-the-loop oversight for safety-critical water-distribution-network applications.
4
The multimodal DataAnalyzer generated valid interpretations of simulation results across system characteristics, dynamics, operations, and scenario simulations.
5
The system coordinates three specialized agents—TaskExecutor, CodeRunner, and DataAnalyzer—to decompose tasks, invoke tools precisely, and manage workflows transparently.
Research Object
water distribution networks (WDNs) modeled and simulated using EPANET
Research Subject
natural language-controlled hydraulic and water-quality simulation, analysis, and workflow management
Publication Details
Publication Date
2026-01-21
Journal
Publisher
ISSN
Cited by
24
Open access PDF
Access Type
Author Information
Download PDF
Subscribe to digest