arxiv: 2604.02876 · v1 · submitted 2026-04-03 · 💻 cs.LG

Toward an Operational GNN-Based Multimesh Surrogate for Fast Flood Forecasting

Valentin Mercier (Toulouse INP , IRIT , EPE UT) , Serge Gratton (IRIT , EPE UT , Toulouse INP) , Lapeyre Corentin (NVIDIA) , Gwena\"el Chevallet This is my paper

Pith reviewed 2026-05-13 19:27 UTC · model grok-4.3

classification 💻 cs.LG

keywords graph neural networksflood forecastingsurrogate modelingmultimesh connectivityhydraulic simulationoperational forecastingTelemac2D

0 comments p. Extension

The pith

A graph neural network surrogate with multimesh connectivity forecasts floods accurately in seconds rather than hours.

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

The paper builds a graph neural network surrogate for fast prediction of floods on the lower Têt River starting from a production Telemac2D model with more than 400,000 nodes. It assembles a database of synthetic yet operationally grounded flood events and trains the surrogate on projected meshes that retain high-fidelity supervision. The work isolates the contributions of an explicit discharge input Q(t), multimesh connectivity that widens the spatial receptive field, and pushforward training for stable rollouts. Among the tested setups, the combination of all three yields the strongest agreement with the reference solver both on hydraulic fields over the surrogate mesh and on inundation maps interpolated to a common 25 m grid. The resulting model produces a 6-hour forecast in roughly 0.4 seconds on one GPU, versus 180 minutes on 56 CPU cores for the original simulation.

Core claim

The combination of Q(t) conditioning, multimesh connectivity, and pushforward training in a GNN-based surrogate provides the best results for emulating high-fidelity 2D flood simulations, with accuracy preserved on both surrogate variables and high-resolution inundation maps while delivering substantial computational speedup.

What carries the argument

The multimesh connectivity in the graph neural network that enlarges the effective spatial receptive field without added depth, together with projected meshes for tractable training and explicit discharge conditioning Q(t).

Load-bearing premise

The collection of synthetic flood events is sufficiently representative of real-world conditions to allow the surrogate to generalize to unseen events.

What would settle it

Comparing the surrogate's 6-hour inundation predictions against observations from an actual flood event outside the training database, or against a new independent high-resolution simulation of such an event.

Figures

Figures reproduced from arXiv: 2604.02876 by EPE UT, EPE UT), Gwena\"el Chevallet, IRIT, Lapeyre Corentin (NVIDIA), Serge Gratton (IRIT, Toulouse INP), Valentin Mercier (Toulouse INP.

**Figure 1.** Figure 1: Overview of the dataset design: (a) spatial mesh-density strategy and (b) synthetic hydrograph families used [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 2.** Figure 2: Distribution of unique edge lengths for the reference Telemac mesh, the projected [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: Ablation of the global discharge feature [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

**Figure 4.** Figure 4: Ablation of multimesh connectivity on the 16 held-out floods, restricted to [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗

**Figure 5.** Figure 5: Ablation of pushforward training on the 16 held-out floods. Left: standard mesh with [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

**Figure 6.** Figure 6: Fine-grid CSI on the 25 m common grid with threshold hε = 5 cm for the four Q(t)-conditioned configurations (E2, E3, E5, E6) on the 16 held-out floods. Solid lines show the mean CSI at each lead time and shaded areas indicate one standard deviation [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗

**Figure 7.** Figure 7: Fine-grid CSI on the 25 m common grid with threshold hε = 30 cm for the four Q(t)-conditioned configurations (E2, E3, E5, E6) on the 16 held-out floods. Solid lines show the mean CSI at each lead time and shaded areas indicate one standard deviation. range of realistic hydrograph shapes and peak discharges. This database constitutes an important contribution of the present work, as it makes it possible to… view at source ↗

**Figure 8.** Figure 8: Qualitative binary inundation map on the [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗

read the original abstract

Operational flood forecasting still relies on high-fidelity two-dimensional hydraulic solvers, but their runtime can be prohibitive for rapid decision support on large urban floodplains. In parallel, AI-based surrogate models have shown strong potential in several areas of computational physics for accelerating otherwise expensive high-fidelity simulations. We address this issue on the lower T\^et River (France), starting from a production-grade Telemac2D model defined on a high-resolution unstructured finite-element mesh with more than $4\times 10^5$ nodes. From this setup, we build a learning-ready database of synthetic but operationally grounded flood events covering several representative hydrograph families and peak discharges. On top of this database, we develop a graph-neural surrogate based on projected meshes and multimesh connectivity. The projected-mesh strategy keeps training tractable while preserving high-fidelity supervision from the original Telemac simulations, and the multimesh construction enlarges the effective spatial receptive field without increasing network depth. We further study the effect of an explicit discharge feature $Q(t)$ and of pushforward training for long autoregressive rollouts. The experiments show that conditioning on $Q(t)$ is essential in this boundary-driven setting, that multimesh connectivity brings additional gains once the model is properly conditioned, and that pushforward further improves rollout stability. Among the tested configurations, the combination of $Q(t)$, multimesh connectivity, and pushforward provides the best overall results. These gains are observed both on hydraulic variables over the surrogate mesh and on inundation maps interpolated onto a common $25\,\mathrm{m}$ regular grid and compared against the original high-resolution Telemac solution. On the studied case, the learned surrogate produces 6-hour predictions in about $0.4\,\mathrm{s}$ on a single NVIDIA A100 GPU, compared with about $180\,\mathrm{min}$ on 56 CPU cores for the reference simulation. These results support graph-based surrogates as practical complements to industrial hydraulic solvers for operational flood mapping.

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.

Desk Editor's Note private letter to a colleague

The GNN surrogate shows solid speedups and component gains on synthetic Têt River floods, but real-event generalization stays untested.

read the letter

The main point is that this paper builds a practical graph neural network surrogate for 2D flood modeling that runs in under a second on GPU while matching the high-fidelity Telemac2D solver on held-out synthetic events. They start from a real production mesh with over 400k nodes, generate a database of representative hydrographs, project to a coarser mesh for training, and add explicit Q(t) conditioning plus multimesh connectivity and pushforward training. The experiments isolate the contributions: Q(t) proves essential for boundary-driven flows, multimesh improves spatial coverage without extra depth, and pushforward helps rollout stability. The best setup delivers usable accuracy on both surrogate-mesh variables and 25 m interpolated inundation maps, with the headline runtime of 0.4 s versus 180 minutes on 56 CPU cores. That combination of projected multimesh, discharge input, and pushforward training is the concrete new piece beyond standard GNN physics applications. The work is grounded in operational constraints and the results are reported clearly enough to see where each change helps. The central limitation is that everything stays inside the synthetic data distribution. No real gauge records, observed inundation extents, or out-of-distribution boundaries are tested, so the assumption that these synthetics are close enough for deployment is not checked. Error distributions and behavior on actual events would be needed to judge operational value. This is useful for groups working on ML surrogates for hydraulic models or operational flood teams looking for fast complements to full solvers. The methods are described at a level that supports follow-up, and the scale of the problem plus the reported gains make it worth sending to peer review for detailed scrutiny on validation and generalization.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces a graph neural network surrogate for accelerating 2D flood simulations on the lower Têt River, starting from a production Telemac2D model with >4e5 nodes. It constructs a database of synthetic but operationally grounded flood events, then trains a GNN on projected meshes with multimesh connectivity, explicit discharge Q(t) conditioning, and pushforward training for autoregressive rollouts. The central claim is that the Q(t)+multimesh+pushforward combination yields the best accuracy on held-out synthetic events, measured both on the surrogate mesh and on 25 m interpolated inundation maps, while delivering a 6-hour forecast in 0.4 s on one A100 GPU versus 180 min on 56 CPU cores for the reference solver.

Significance. If the reported accuracy and generalization hold beyond the synthetic database, the work provides a practical route to real-time operational flood mapping that could complement existing high-fidelity solvers. The multimesh connectivity strategy and the controlled ablation of Q(t), connectivity, and pushforward training constitute a clear technical contribution to GNN-based surrogates for boundary-driven environmental physics problems.

major comments (2)

[Experiments and Results] Experiments / Results: All quantitative comparisons (including the superiority of the full Q(t)+multimesh+pushforward configuration) are performed exclusively on held-out synthetic hydrographs generated by the identical Telemac2D model used to create the training set. No real gauge data, observed inundation extents, or out-of-distribution boundary conditions are evaluated, leaving the transfer to operational real-world events untested and therefore weakening support for the abstract's claim of practical complementarity to industrial solvers.
[Abstract and Introduction] Abstract and Introduction: The manuscript asserts that the surrogate 'supports graph-based surrogates as practical complements ... for operational flood mapping,' yet the only evidence consists of performance on synthetic events whose distributional similarity to real floods is asserted but not stress-tested; a concrete discussion or additional experiment addressing this transfer is required to substantiate the central operational claim.

minor comments (1)

[Methods] Methods: The exact GNN architecture hyperparameters, training schedule, and precise definition of the multimesh projection operator should be stated explicitly (or supplied as supplementary material) to enable full reproducibility.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive feedback and for recognizing the technical contributions of the multimesh GNN surrogate. We address the two major comments below, clarifying the scope of our synthetic but operationally grounded evaluation while proposing revisions to better contextualize the path to real-world use.

read point-by-point responses

Referee: [Experiments and Results] Experiments / Results: All quantitative comparisons (including the superiority of the full Q(t)+multimesh+pushforward configuration) are performed exclusively on held-out synthetic hydrographs generated by the identical Telemac2D model used to create the training set. No real gauge data, observed inundation extents, or out-of-distribution boundary conditions are evaluated, leaving the transfer to operational real-world events untested and therefore weakening support for the abstract's claim of practical complementarity to industrial solvers.

Authors: We agree that all reported metrics are computed on held-out synthetic hydrographs. These events were constructed from representative hydrograph families and peak discharges drawn from historical gauge records on the Têt River, providing a controlled yet relevant testbed that isolates the impact of Q(t) conditioning, multimesh connectivity, and pushforward training. Direct comparison to the reference Telemac2D solution on the same events enables precise error quantification that would be impossible with sparse real observations. We will add a new subsection titled 'Path to Operational Deployment' in the Discussion that explicitly discusses distributional similarity to real floods, potential domain shifts, and future strategies such as fine-tuning on gauge data. No new real-data experiments are added, as they fall outside the current database. revision: partial
Referee: [Abstract and Introduction] Abstract and Introduction: The manuscript asserts that the surrogate 'supports graph-based surrogates as practical complements ... for operational flood mapping,' yet the only evidence consists of performance on synthetic events whose distributional similarity to real floods is asserted but not stress-tested; a concrete discussion or additional experiment addressing this transfer is required to substantiate the central operational claim.

Authors: We will revise the abstract to state that the results 'support graph-based surrogates as practical complements to industrial hydraulic solvers for operational flood mapping on synthetic but representative events.' Parallel adjustments will be made in the introduction and conclusions. We will also insert a concise paragraph in the Discussion that outlines the transfer challenges and concrete next steps (e.g., assimilation of real gauge time series), thereby providing the requested discussion without overstating current evidence. revision: yes

standing simulated objections not resolved

Direct quantitative validation against observed real-world inundation extents or out-of-distribution gauge events, which would require additional observational datasets not present in the current synthetic database.

Circularity Check

0 steps flagged

No significant circularity; derivation relies on independent high-fidelity data

full rationale

The paper constructs a GNN surrogate by training on synthetic flood events generated by the external Telemac2D solver and evaluates performance on held-out synthetic hydrographs using direct comparisons to the original solver outputs on both surrogate-mesh variables and interpolated 25 m grids. No equations reduce reported predictions to fitted quantities by construction, no self-citations serve as load-bearing uniqueness theorems, and no ansatzes or renamings collapse the central claims (conditioning on Q(t), multimesh connectivity, pushforward) into the inputs themselves. The approach remains self-contained against the independent reference solver.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on the assumption that mesh projection and multimesh connectivity preserve the essential hydraulic dynamics while making training feasible, plus the representativeness of the synthetic event database.

free parameters (2)

GNN architecture hyperparameters
Network depth, width, and aggregation choices are selected to balance accuracy and training cost on the projected meshes.
Training schedule parameters
Learning rate, rollout length, and pushforward frequency are tuned on the synthetic flood database.

axioms (2)

domain assumption Projected meshes retain sufficient hydraulic information from the original >4e5-node Telemac mesh for accurate surrogate learning.
Invoked to justify keeping training tractable while using high-fidelity supervision.
domain assumption The synthetic database covers representative hydrograph families and peak discharges for operational conditions.
Used to generate the training data from the production-grade model.

pith-pipeline@v0.9.0 · 5708 in / 1476 out tokens · 36020 ms · 2026-05-13T19:27:11.905581+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

[1]

Statista

Economic insured losses caused by floods worldwide from 2013 to 2023. Statista. https://www.statista. com/statistics/1326526/economic-losses-floods-worldwide-by-type/ . Accessed 2025

work page arXiv 2013
[2]

Hec-ras river analysis system

Gary W Brunner. Hec-ras river analysis system. hydraulic reference manual. version 1.0. 1997

work page 1997
[3]

Hydrodynamics of free surface flows: modelling with the finite element method

Jean-Michel Hervouet. Hydrodynamics of free surface flows: modelling with the finite element method . John Wiley & Sons, 2007

work page 2007
[4]

Deep learning methods for flood mapping: a review of existing applications and future research directions

Roberto Bentivoglio, Elvin Isufi, Sebastian Nicolaas Jonkman, and Riccardo Taormina. Deep learning methods for flood mapping: a review of existing applications and future research directions. Hydrology and Earth System Sciences Discussions, 2022:1–50, 2022

work page 2022
[5]

An ann-based emulation modelling framework for flood inundation modelling: Application, challenges and future directions.Environmental Modelling & Software, 124:104587, 2020

Haibo Chu, Wenyan Wu, Quan J Wang, Rory Nathan, and Jiahua Wei. An ann-based emulation modelling framework for flood inundation modelling: Application, challenges and future directions.Environmental Modelling & Software, 124:104587, 2020

work page 2020
[6]

Fusing stacked autoencoder and long short-term memory for regional multistep-ahead flood inundation forecasts

I-Feng Kao, Jia-Yi Liou, Meng-Hsin Lee, and Fi-John Chang. Fusing stacked autoencoder and long short-term memory for regional multistep-ahead flood inundation forecasts. Journal of Hydrology, 598:126371, 2021

work page 2021
[7]

Real-time flood maps forecasting for dam-break scenarios with a transformer-based deep learning model

Matteo Pianforini, Susanna Dazzi, Andrea Pilzer, and Renato Vacondio. Real-time flood maps forecasting for dam-break scenarios with a transformer-based deep learning model. Journal of Hydrology, 635:131169, 2024

work page 2024
[8]

Large-scale flood modeling and forecasting with floodcast

Qingsong Xu, Yilei Shi, Jonathan L Bamber, Chaojun Ouyang, and Xiao Xiang Zhu. Large-scale flood modeling and forecasting with floodcast. Water Research, 264:122162, 2024

work page 2024
[9]

Pfaff, M

Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter W Battaglia. Learning mesh-based simulation with graph networks. arXiv preprint arXiv:2010.03409, 2020. 13 A PREPRINT - APRIL 6, 2026

work page arXiv 2010
[10]

Rapid spatio-temporal flood modelling via hydraulics-based graph neural networks

Roberto Bentivoglio, Elvin Isufi, Sebastiaan Nicolas Jonkman, and Riccardo Taormina. Rapid spatio-temporal flood modelling via hydraulics-based graph neural networks. Hydrology and Earth System Sciences, 27(23):4227– 4246, 2023

work page 2023
[11]

Multi-scale hydraulic graph neural networks for flood modelling

Roberto Bentivoglio, Elvin Isufi, Sebastiaan Nicolas Jonkman, and Riccardo Taormina. Multi-scale hydraulic graph neural networks for flood modelling. EGUsphere, 2024:1–28, 2024

work page 2024
[12]

In- terpretable physics-informed graph neural networks for flood forecasting.Computer-Aided Civil and Infrastructure Engineering, 40:2629–2649, 2025

Mehdi Taghizadeh, Majid Shafiee-Jood, Zanko Zandsalimi, Negin Alemazkoor, and Mohammad Amin Nabian. In- terpretable physics-informed graph neural networks for flood forecasting.Computer-Aided Civil and Infrastructure Engineering, 40:2629–2649, 2025

work page 2025
[13]

The finite volume method in computational fluid dynamics

F Moukalled L Mangani M Darwish. The finite volume method in computational fluid dynamics. 2016

work page 2016
[14]

Finite element procedures

Klaus-Jürgen Bathe. Finite element procedures. Klaus-Jurgen Bathe, 2006

work page 2006
[15]

Multiscale meshgraphnets, 2022

Meire Fortunato, Tobias Pfaff, Peter Wirnsberger, Alexander Pritzel, and Peter Battaglia. Multiscale meshgraphnets, 2022

work page 2022
[16]

Message passing neural pde solvers, 2022

Johannes Brandstetter, Daniel Worrall, and Max Welling. Message passing neural pde solvers, 2022. 14

work page 2022