pith. machine review for the scientific record. sign in

arxiv: 2602.07299 · v2 · submitted 2026-02-07 · 💻 cs.NE · cs.SY· eess.SY

Recognition: 1 theorem link

· Lean Theorem

Optimizing Chlorination in Water Distribution Systems via Surrogate-assisted Neuroevolution

Rivaaj Monsia , Daniel Young , Olivier Francon , Risto Miikkulainen
Authors on Pith no claims yet

Pith reviewed 2026-05-16 07:00 UTC · model grok-4.3

classification 💻 cs.NE cs.SYeess.SY
keywords neuroevolutionwater distribution systemschlorine injectionsurrogate modelingmulti-objective optimizationNEATNSGA-IIreinforcement learning
0
0 comments X

The pith

Neuroevolution with a neural surrogate produces practical Pareto-optimal chlorine injection policies for water distribution systems that outperform PPO.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that evolving neural network controllers with NEAT, combined with NSGA-II multi-objective optimization and a neural network surrogate trained to emulate the EPANET simulator, can generate effective strategies for maintaining disinfectant levels in large, complex water networks. Traditional algorithms struggle with the nonlinear fluid dynamics, so this framework optimizes four goals at once: minimizing total chlorine use, achieving uniform concentrations across the network, respecting safe maximum limits, and spacing injections regularly in time. The evolved policies form a diverse set of trade-off solutions that are ready for practical use and surpass standard reinforcement learning with PPO. A sympathetic reader would care because the approach reduces computational barriers for controlling real urban water infrastructure while improving safety and efficiency.

Core claim

Neural networks evolved with NEAT and optimized by NSGA-II on four objectives while evaluated against a neural network surrogate of EPANET yield a diverse range of Pareto-optimal chlorination policies that can be implemented in practice and that outperform PPO.

What carries the argument

Surrogate-assisted neuroevolution, where NEAT evolves neural controllers for timed chlorine injections at network locations and NSGA-II searches for trade-offs across objectives using a neural network trained to approximate the EPANET hydraulic simulator.

Load-bearing premise

The neural network surrogate accurately reproduces the relevant dynamics of the EPANET simulator for the conditions encountered during optimization.

What would settle it

Running the evolved policies on the full EPANET simulator or a physical water system and measuring whether chlorine concentrations remain homogeneous, within safe bounds, and with the expected total injection amounts.

Figures

Figures reproduced from arXiv: 2602.07299 by Daniel Young, Olivier Francon, Risto Miikkulainen, Rivaaj Monsia.

Figure 1
Figure 1. Figure 1: Topology of the water distribution network utilized [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The outer loop of the Evolutionary Surrogate [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of 𝑇𝜙 vs. 𝑆𝜃 vs. true values for 𝑜𝑡+1/𝑜ˆ𝑡+1. Low error in 𝑆𝜃 predictions despite large variations in con￾centration suggests that the distillation from 𝑇𝜙 is reliable, and the resulting surrogate model can be used to optimize the controller. The results in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Pairwise population fronts with (a) two, (b) three, [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: NEAT innovation resulting from fine-tuning the [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of 𝑇𝜙 vs. 𝑆𝜃 vs. True values for a) scaled Δ𝑡 and b) unscaled Δ𝑡 . Low error in 𝑆𝜃 predictions indicates that the composition of scaling with incremental predictions enables accurate modeling of the hydraulic state. those further apart (6-16) vary significantly. In principle, it may be possible to normalize the outputs of each sensor, but this variation across sensors would still remain: there a… view at source ↗
Figure 9
Figure 9. Figure 9: Configuration file for NEAT experiments via NSGA-II using [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
read the original abstract

Ensuring the microbiological safety of large, heterogeneous water distribution systems (WDS) typically requires managing appropriate levels of disinfectant residuals including chlorine. WDS include complex fluid interactions that are nonlinear and noisy, making such maintenance a challenging problem for traditional control algorithms. This paper proposes an evolutionary framework to this problem based on neuroevolution, multi-objective optimization, and surrogate modeling. Neural networks were evolved with NEAT to inject chlorine at strategic locations in the distribution network at select times. NSGA-II was employed to optimize four objectives: minimizing the total amount of chlorine injected, keeping chlorine concentrations homogeneous across the network, ensuring that maximum concentrations did not exceed safe bounds, and distributing the injections regularly over time. Each network was evaluated against a surrogate model, i.e.\ a neural network trained to emulate EPANET, an industry-level hydraulic WDS simulator that is accurate but infeasible in terms of computational cost to support machine learning. The evolved controllers produced a diverse range of Pareto-optimal policies that could be implemented in practice, outperforming PPO, a standard reinforcement learning method. The results thus suggest a pathway toward improving urban water systems, and highlight the potential of using evolution with surrogate modeling to optimize complex real-world systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes a surrogate-assisted neuroevolution framework for optimizing chlorine injection controllers in water distribution systems. Neural networks are evolved via NEAT and optimized with NSGA-II across four objectives (minimize total chlorine injected, maximize homogeneity of concentrations, respect safe maximum bounds, and regularize injection timing over time). Evaluations use a neural network surrogate trained to emulate the EPANET simulator. The central claim is that the resulting Pareto-optimal policies are diverse, practical to implement, and outperform PPO.

Significance. If the surrogate validation holds, the work demonstrates a practical pathway for applying neuroevolution and multi-objective surrogate modeling to computationally expensive real-world control problems in infrastructure, with potential impact on urban water safety and efficiency.

major comments (2)
  1. [Surrogate Model] Surrogate validation section: no test-set error, coverage metrics, or hold-out validation is reported for the surrogate NN specifically on trajectories or policies generated during NSGA-II search. This directly undermines the central empirical claim of outperformance over PPO and transferability to the physical system, as objective values could arise from surrogate exploitation rather than EPANET dynamics.
  2. [Experimental Results] Results section: the manuscript supplies no quantitative metrics (e.g., objective values, number of runs, statistical tests), details on objective normalization/weighting, or surrogate validation error, preventing verification of the Pareto front diversity and PPO outperformance claims.
minor comments (2)
  1. [Abstract] Abstract and methods: expand on surrogate NN architecture, training dataset size, and validation split to support reproducibility.
  2. [Methods] Notation and figures: clarify how the four objectives are scaled or combined in the NSGA-II fitness function.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the constructive referee report. We address each major comment below and will revise the manuscript to incorporate additional validation and quantitative details as requested.

read point-by-point responses

  1. Referee: [Surrogate Model] Surrogate validation section: no test-set error, coverage metrics, or hold-out validation is reported for the surrogate NN specifically on trajectories or policies generated during NSGA-II search. This directly undermines the central empirical claim of outperformance over PPO and transferability to the physical system, as objective values could arise from surrogate exploitation rather than EPANET dynamics.

    Authors: We agree that the surrogate validation reporting is insufficient as presented. The manuscript describes training the surrogate on EPANET-generated data but does not include the requested quantitative metrics or hold-out evaluation on NSGA-II trajectories. In revision we will add a dedicated validation subsection reporting test-set error, coverage metrics, and hold-out performance specifically on policies sampled during the search. This will directly address concerns about surrogate exploitation versus true EPANET dynamics. revision: yes

  2. Referee: [Experimental Results] Results section: the manuscript supplies no quantitative metrics (e.g., objective values, number of runs, statistical tests), details on objective normalization/weighting, or surrogate validation error, preventing verification of the Pareto front diversity and PPO outperformance claims.

    Authors: We acknowledge that the results section lacks the quantitative detail needed for verification. The current text emphasizes qualitative diversity and outperformance without reporting specific objective values, run counts, statistical comparisons, normalization procedures, or surrogate error. We will expand the section in revision to supply these elements, including objective values for representative policies, details on the experimental protocol, and explicit normalization/weighting information. revision: yes

Circularity Check

0 steps flagged

Derivation chain self-contained with no reductions to inputs by construction

full rationale

The paper trains a surrogate neural network separately to emulate EPANET, then applies NEAT neuroevolution and NSGA-II multi-objective optimization to evolve controllers whose performance is measured on that surrogate. The outperformance claim versus PPO is a direct comparison on the identical evaluation setup. No equations, fitted parameters, or self-citations reduce the reported objective values or Pareto policies back to the surrogate training data or evolutionary inputs by definition. The surrogate training step and the subsequent search are distinct external processes, so the derivation remains non-circular.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The framework rests on the domain assumption that EPANET provides a sufficiently accurate reference model and that the surrogate neural network can be trained to emulate it with acceptable error for optimization purposes.

axioms (1)
  • domain assumption EPANET simulator accurately models hydraulic and chlorine transport behavior in the target networks
    The surrogate is trained to emulate EPANET, which is treated as the ground-truth reference.

pith-pipeline@v0.9.0 · 5525 in / 1232 out tokens · 50182 ms · 2026-05-16T07:00:32.301684+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages · 3 internal anchors

  1. [1]

    Environmental Protection Agency. [n. d.]. https://www.epa.gov/water-research/ epanet

    work page

  2. [2]

    Kyriakou, Stelios G

    André Artelt, Marios S. Kyriakou, Stelios G. Vrachimis, Demetrios G. Eliades, Barbara Hammer, and Marios M. Polycarpou. 2024. EPyT-Flow: A Toolkit for Generating Water Distribution Network Data.Journal of Open Source Software9, 103 (2024), 7104. doi:10.21105/joss.07104

    work page doi:10.21105/joss.07104 2024

  3. [3]

    Vrachimis, Demetrios G

    André Artelt, Janine Strotherm Luca Hermes, Barbara Hammer, Stelios G. Vrachimis, Demetrios G. Eliades Marios S. Kyriakou, Marios M. Polycarpou, Sotirios Paraskevopoulos, Stefanos Vrochidis, Riccardo Taormina, Dragan Savic, and Phoebe Koundouri. 2025. 1st AI for Drinking Water Chlorination Challenge

    work page 2025

  4. [4]

    Inaam Ashraf, Janine Strotherm, Luca Hermes, and Barbara Hammer. 2024. Physics-informed graph neural networks for water distribution systems.Pro- ceedings of the AAAI Conference on Artificial Intelligence38, 20 (Mar 2024), 21905–21913. doi:10.1609/aaai.v38i20.30192

    work page doi:10.1609/aaai.v38i20.30192 2024

  5. [5]

    Pramod R. Bhave. 1991.Analysis of flow in water distribution networks. Technomic Pub. Co

    work page 1991

  6. [6]

    Emily Clements, Katherine Crank, Robert Nerenberg, Ariel Atkinson, Daniel Gerrity, and Deena Hannoun. 2024. Quantitative microbial risk assessment framework incorporating water ages with legionella pneumophila growth rates. Environmental Science & Technology58, 15 (Apr 2024), 6540–6551. doi:10.1021/ acs.est.4c01208

    work page 2024

  7. [7]

    2015–2025

    LLC CodeReclaimers. 2015–2025. NEAT-Python. urlhttps://github.com/CodeReclaimers/neat-python

    work page 2015

  8. [8]

    Creaco, A

    E. Creaco, A. Campisano, N. Fontana, G. Marini, P.R. Page, and T. Walski. 2019. Real time control of water distribution networks: A state-of-the-art review.Water Research161 (Sep 2019), 517–530. doi:10.1016/j.watres.2019.06.025

    work page doi:10.1016/j.watres.2019.06.025 2019

  9. [9]

    K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan. 2002. A fast and elitist multiobjec- tive genetic algorithm: NSGA-II.IEEE Transactions on Evolutionary Computation 6, 2 (2002), 182–197. doi:10.1109/4235.996017

    work page doi:10.1109/4235.996017 2002

  10. [10]

    Elsherif, Ahmad F

    Salma M. Elsherif, Ahmad F. Taha, and Ahmed A. Abokifa. 2024. Disinfectant control in drinking water networks: Integrating advection–dispersion–reaction models and byproduct constraints.Water Research267 (Dec 2024), 122441. doi:10. 1016/j.watres.2024.122441

    work page arXiv 2024

  11. [11]

    Xudong Fan, Xijin Zhang, and Xiong Yu. 2021. Machine learning model and strategy for fast and accurate detection of leaks in Water Supply Network.Journal of Infrastructure Preservation and Resilience2, 1 (Apr 2021). doi:10.1186/s43065- 021-00021-6

    work page doi:10.1186/s43065- 2021

  12. [12]

    Ian Fisher, George Kastl, and Arumugam Sathasivan. 2011. Evaluation of suitable chlorine bulk-decay models for water distribution systems.Water Research45, 16 (Oct 2011), 4896–4908. doi:10.1016/j.watres.2011.06.032

    work page doi:10.1016/j.watres.2011.06.032 2011

  13. [13]

    Olivier Francon, Santiago Gonzalez, Babak Hodjat, Elliot Meyerson, Risto Mi- ikkulainen, Xin Qiu, and Hormoz Shahrzad. 2020. Effective reinforcement learning through evolutionary surrogate-assisted prescription.Proceedings of the 2020 Genetic and Evolutionary Computation Conference(Jun 2020), 814–822. doi:10.1145/3377930.3389842

    work page doi:10.1145/3377930.3389842 2020

  14. [14]

    Walter M. Grayman. 2008. A Quarter of a Century of Water Quality Modeling in Distribution Systems.Water Distribution Systems Analysis Symposium 2006(Mar 2008), 1–12. doi:10.1061/40941(247)4

    work page doi:10.1061/40941(247)4 2008

  15. [15]

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the Knowledge in a Neural Network. arXiv:1503.02531 [stat.ML] https://arxiv.org/abs/1503.02531

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  16. [16]

    Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long Short-Term Memory. Neural Computation9, 8 (1997), 1735–1780. doi:10.1162/neco.1997.9.8.1735

    work page doi:10.1162/neco.1997.9.8.1735 1997

  17. [17]

    Pei Hua, Ekaterina Vasyukova, and Wolfgang Uhl. 2015. A variable reaction rate model for chlorine decay in drinking water due to the reaction with dissolved organic matter.Water Research75 (May 2015), 109–122. doi:10.1016/j.watres. 2015.01.037

    work page doi:10.1016/j.watres 2015

  18. [18]

    Zhiheng Huang, Wei Xu, and Kai Yu. 2015. Bidirectional LSTM-CRF Models for Sequence Tagging. arXiv:1508.01991 [cs.CL] https://arxiv.org/abs/1508.01991

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  19. [19]

    Sinan Ibrahim, Mostafa Mostafa, Ali Jnadi, Hadi Salloum, and Pavel Osinenko

    work page

  20. [20]

    arXiv:2408.10215 [cs.LG] https: //arxiv.org/abs/2408.10215

    Comprehensive Overview of Reward Engineering and Shaping in Ad- vancing Reinforcement Learning Applications. arXiv:2408.10215 [cs.LG] https: //arxiv.org/abs/2408.10215

    work page arXiv

  21. [21]

    Indrajit Kalita, Andreas Kamilaris, Paul Havinga, and Igor Reva. 2024. Assessing the health impact of disinfection byproducts in drinking water.ACS ES&T Water 4, 4 (Mar 2024), 1564–1578. doi:10.1021/acsestwater.3c00664

    work page doi:10.1021/acsestwater.3c00664 2024

  22. [22]

    Taehyeon Kim, Jaehoon Oh, Nak Yil Kim, Sangwook Cho, and Se-Young Yun

    work page

  23. [23]

    doi:10.24963/ijcai.2021/362

    Comparing Kullback-Leibler divergence and mean squared error loss in knowledge distillation.Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence(Aug 2021), 2628–2635. doi:10.24963/ijcai.2021/362

    work page doi:10.24963/ijcai.2021/362 2021

  24. [24]

    Kyriakou, Marios Demetriades, Stelios G

    Marios S. Kyriakou, Marios Demetriades, Stelios G. Vrachimis, Demetrios G. Eliades, and Marios M. Polycarpou. 2023. EPyT: An EPANET-python toolkit for Smart Water Network Simulations.Journal of Open Source Software8, 92 (Dec 2023), 5947. doi:10.21105/joss.05947

    work page doi:10.21105/joss.05947 2023

  25. [25]

    Jing Li, Cuimin Feng, Ying Li, Weiqi Yang, and Zibin Zhang. 2022. Formation and influencing factors of disinfection by-products from bacterial materials in drinking water distribution systems.Water Supply22, 9 (Sep 2022), 7319–7336. doi:10.2166/ws.2022.319

    work page doi:10.2166/ws.2022.319 2022

  26. [26]

    United Nations. [n. d.]. https://sdgs.un.org/goals

    work page

  27. [27]

    Roth and David A

    Damon K. Roth and David A. Cornwell. 2018. DBP impacts from increased chlorine residual requirements.Journal A WW A110, 2 (Feb 2018), 13–28. doi:10. 5942/jawwa.2018.110.0004

    work page 2018

  28. [28]

    Iqbal H. Sarker. 2021. Machine learning: Algorithms, real-world applications and Research Directions.SN Computer Science2, 3 (Mar 2021). doi:10.1007/s42979- 021-00592-x

    work page doi:10.1007/s42979- 2021

  29. [29]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov

    work page

  30. [30]

    Proximal Policy Optimization Algorithms

    Proximal Policy Optimization Algorithms.CoRRabs/1707.06347 (2017). arXiv:1707.06347 http://arxiv.org/abs/1707.06347

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  31. [31]

    Julius Schöning, Adrian Riechmann, and Hans-Jürgen Pfisterer. 2022. AI for closed-loop control systems.2022 14th International Conference on Machine Learning and Computing (ICMLC)(Feb 2022), 318–323. doi:10.1145/3529836. 3529952

    work page doi:10.1145/3529836 2022

  32. [32]

    Seto, Michail Fragkias, Burak Güneralp, and Michael K

    Karen C. Seto, Michail Fragkias, Burak Güneralp, and Michael K. Reilly. 2011. A Meta-Analysis of Global Urban Land Expansion.PLoS ONE6, 8 (Aug 2011). doi:10.1371/journal.pone.0023777

    work page doi:10.1371/journal.pone.0023777 2011

  33. [33]

    Aamir Mazhar, Sirajuddin Ahmed, Beni Lew, and Nadeem Khalil

    Mehreen Shah, Mohd. Aamir Mazhar, Sirajuddin Ahmed, Beni Lew, and Nadeem Khalil. 2024. Recent trends in controlling the disinfection by-products before their formation in drinking water: A Review.Drinking Water Disinfection By-products (2024), 177–192. doi:10.1007/978-3-031-49047-7_9

    work page doi:10.1007/978-3-031-49047-7_9 2024

  34. [34]

    Petru Soviany, Radu Tudor Ionescu, Paolo Rota, and Nicu Sebe. 2022. Curriculum Learning: A Survey. arXiv:2101.10382 [cs.LG] https://arxiv.org/abs/2101.10382

    work page arXiv 2022

  35. [35]

    Kenneth O Stanley and Risto Miikkulainen. 2002. Evolving neural networks through augmenting topologies.Evolutionary computation10, 2 (2002), 99–127

    work page 2002

  36. [36]

    Junfeng Tang, Handing Wang, and Lin Xiong. 2023. Surrogate-assisted multi- objective optimization via knee-oriented pareto front estimation.Swarm and Evolutionary Computation77 (Mar 2023), 101252. doi:10.1016/j.swevo.2023.101252

    work page doi:10.1016/j.swevo.2023.101252 2023

  37. [37]

    Yipeng Wu, Xiaoting Wang, Shuming Liu, Xipeng Yu, and Xue Wu. 2023. A weighting strategy to improve water demand forecasting performance based on spatial correlation between multiple sensors.Sustainable Cities and Society93 (Jun 2023), 104545. doi:10.1016/j.scs.2023.104545

    work page doi:10.1016/j.scs.2023.104545 2023

  38. [38]

    Qingwei Zhou, Zhengfu Bian, Dejun Yang, and Li Fu. 2023. Stability of drinking water distribution systems and control of disinfection by-products.Toxics11, 7 (Jul 2023), 606. doi:10.3390/toxics11070606 GECCO ’26, July 13–17, 2026, San Jose, Costa Rica Monsia, et. al. A Reward Engineering To evolve agents with NSGA-II, it is necessary to quantify the perfo...

    work page doi:10.3390/toxics11070606 2023