Scalable and Robust Physics-Informed Graph Neural Networks for Water Distribution Systems
Inaam Ashraf, André Artelt, Barbara Hammer · 2025
Read Scalable and Robust Physics-Informed Graph Neural Networks for Water Distribution Systems by Inaam Ashraf, André Artelt, Barbara Hammer (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-12164.txt.
Introduction
Inaam Ashraf, André Artelt, and Barbara Hammer built a graph neural network in 2025 that simulates city water networks fast, without forgetting the physics. Water distribution systems are critical infrastructure under pressure from climate change and growing cities. Planners need many simulations to expand and repair them, and the accurate simulators are slow.
Their surrogate pairs an improved graph architecture with a physics-informed training method and a normalization that keeps the physics intact. An innovative training scheme hardens the model for demands and pipe sizes outside its training range. On several real networks it beat the best current learning model, and it scales to larger, more realistic systems than before. The catalog creator field appends a funding note; the byline governs. Readers get a simulator stand-in aimed at closing the gap between model water and the real pipes.
Catalog record
| CREATOR | Inaam Ashraf, André Artelt, Barbara Hammer |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 41 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-12164
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Scalable and Robust Physics-Informed Graph Neural Networks for Water Distribution Systems Thanks: We gratefully acknowledge funding from the European Research Council (ERC) under the ERC Synergy Grant Water-Futures (Grant agreement No. 951424). Inaam Ashraf, André Artelt, Barbara Hammer Abstract Water distribution systems (WDSs) are an important part of critical infrastructure becoming increasingly significant in the face of climate change and urban population growth. We propose a robust and scalable surrogate deep learning (DL) model to enable efficient planning, expansion, and rehabilitation of WDSs. Our approach incorporates an improved graph neural network architecture, an adapted physics-informed algorithm, an innovative training scheme, and a physics-preserving data normalization method. Evaluation results on a number of WDSs demonstrate that our model outperforms the current state-of-the-art DL model. Moreover, our method allows us to scale the model to bigger and more realistic WDSs. Furthermore, our approach makes the model more robust to out-of-distribution input features (demands, pipe diameters). Hence, our proposed method constitutes a significant step towards bridging the simulation-to-real gap in the use of artificial intelligence for WDSs. I Introduction Reliable and robust water distribution systems (WDSs) are of utmost importance for a sustainable urban society. It is predicted that by the year 2050, more than two-thirds of the world’s population will live in cities [ 1 ] . Since WDSs are responsible for providing clean water to all, their significance aligns directly with UN’s sustainable developmental goal (SDG) 6 – “Clean water and sanitation for all” . Artificial intelligence (AI) has great potential for developing surrogate models and digital twins for revamping the planning, monitoring, and rehabilitation of critical infrastructure [ 2 , 3 , 4 ] . Some work has already been done in this regard [ 5 , 6 ] . In this work, we reduce the gap between simulation and the real world by proposing a robust surrogate DL model that can directly be used for planning, expansion, and rehabilitation of WDSs. A water distribution system is a network of N_{n} nodes connected through N_{e} links. Consumer nodes are called junctions, large water sources like lakes and rivers are reservoirs, and artificial water sources are water tanks. The majority of links in a WDS are water pipes. At some locations, different kinds of valves and pumps are used to maintain water pressure. While planning or expanding a WDS, the structure of the network needs to be designed. For rehabilitation, the task is often to find the optimal changes required to the structure or physical components of the WDS given certain constraints in the future. In both cases, the most important task is to estimate the state of the WDS. This state refers to values of pressure head at every node and water flow through every link given the demands at every consumer node and the pressure heads at the reservoirs. EPANET is the state-of-the-art (SOTA) hydraulic simulator that solves a system of equations to solve the task of hydraulic state estimation in WDSs [ 7 ] . It is highly accurate but takes considerable time when numerous simulations are required, which is often the case in the planning, expansion, and rehabilitation of WDSs. Recently, [ 8 ] proposed a physics-informed DL surrogate model for the hydraulic state estimation showing promising results. However, their approach has limited generalizability to unseen input features. Moreover, since they only experimented with rather small WDSs, it is unclear how well their model scales to larger WDSs. We build on their approach and provide the following key contributions: • We propose a DL surrogate model incorporating an improved GNN architecture, an adapted physics-informed algorithm, an innovative training scheme, and a physics-preserving data normalization method. • We implement two complex types of links i.e. pump and pressure-reducing valve (PRV) to allow the application of the model to realistic WDSs. • We evaluate our method on a number of WDSs showing great accuracy and efficiency gains. • Our model performs equally well on both small and larger realistic WDSs and is robust to both unseen in-distribution and out-of-distribution input features as demonstrated by extensive experiments. II Related Work AI and more specifically machine learning (ML) has been used to solve certain tasks in WDS. These include demand prediction [ 9 ] , leakage detection [ 10 , 11 ] and sensor fault detection [ 12 ] . Moreover, there exist some approaches for the tasks like optimal sensor placement [ 13 , 14 ] . In the real world, pressure readings from a few sensors are available. Hence, the task of estimating the state of the WDS at every node using these sparse readings is important for leakage and sensor fault detection. This task has attracted considerable attention recently as demonstrated by [ 14 , 15 , 16 , 17 , 18 ] . Graph neural networks (GNNs) are a promising DL method to deal with network structures like WDS. There exist a lot of different GNN architectures such as spectral graph convolutional neural networks (GCNs) [ 19 , 20 , 21 , 22 , 23 , 24 ] , spatial GCNs [ 25 , 26 , 27 , 28 , 29 , 30 ] , and recursive graph and tree models [ 31 , 32 ] . Similar to [ 8 ] , our proposed architecture employs custom message-passing GNNs that fall under the umbrella of spatial GCNs. The task of hydraulic state estimation in WDS involves estimating the pressure head at every node and water flow through every link given the demands at every consumer node and the pressure heads at the reservoirs. EPANET is the prevalent hydraulic simulator that solves a system of equations to solve this task [ 7 ] . A faster alternative method using edge diffusion was recently proposed by [ 33 ] with promising results. However, it is unclear how well the method generalizes to changes in input features. A few DL surrogate models for hydraulic state estimation already exist. Reference [ 18 ] combines the inputs required by EPANET with sparse sensor measurements to solve this task. However, the efficacy of their method is indeterminate given their limited empirical evaluation. Using the same input data as [ 18 ] , [ 34 ] proposed surrogate metamodels using edge-based GNNs. Although they demonstrate good results, their approach is limited to only single-reservoir WDSs. Reference [ 8 ] are the first to solve the task of state estimation in WDS using a DL model showing promising results. However, their approach suffers from limited generalizability to unseen input features (demands, pipe diameters, etc.). Moreover, since they only experimented with rather small WDSs, it is unclear how well their model scales to larger and more realistic WDSs. WDSs are characterized by different features (pressures, demands, flows, pipe attributes, etc.) having values of different magnitudes that are tied together through physical laws. Other domains in critical infrastructure like electricity grids also exhibit similar characteristics [ 35 ] . Hence, traditional data normalization methods used in DL cannot be directly applied without violating the physical relationships. This constitutes an open problem that is addressed in this paper. III Methodology We propose a DL surrogate model comprising an improved GNN architecture, an adapted physics-informed algorithm incorporating pumps