the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 08 Jul 2026
High-resolution rain-on-grid hydrodynamic modelling can replace hydrological models for catchment-scale flood simulation: the case of the 2021 Ahr catchment flood
Abstract. Flood simulation and forecasting in mid- to large-sized catchments has long relied on coupling hydrological and hydrodynamic models, a two-tiered modelling strategy driven mainly by the assumption that physically-based hydrodynamic simulations at the resolutions needed to resolve channels and floodplains are computationally prohibitive at catchment scale. Recent advances in multi-GPU high-performance computing, together with increasing availability of high-resolution geospatial data are challenging this assumption. In this study, we test whether a fully hydrodynamic, rain-on-grid approach can be a standalone alternative to the traditional coupled chain, using the July 2021 flood in the Ahr catchment, Germany. We apply the performance-portable multi-GPU shallow-water solver SERGHEI to the entire ~900 km2 Ahr catchment at dx = 2, 5, and 10 m, forcing it directly with the 5-minute RADOLAN precipitation product. Assessed against the observed event, the rain-on-grid setup reproduces the flood characteristics remarkably well. Model skill metrics are comparable to –and in some metrics improving–, those reported by previous inundation studies relying on reconstructed hydrographs or hydrological-hydrodynamic two-tiered modelling chains for the same event. Remarkably the present setup using no hydrological model, no reconstructed inflow hydrograph, and no parameter calibration, and covering a domain roughly an order of magnitude larger, achieves very good results. We further show that the approach scales efficiently across two generations of GPU supercomputers, completing the 96-hour event between roughly 18 and 420 times faster than real time, well within operational early-warning requirements. The results suggest that, for catchments of this scale and event types of this nature, catchment-scale rain-on-grid hydrodynamic modelling has reached a level of maturity that justifies its use as a standalone alternative to the more traditional coupled hydrological–hydrodynamic chain.
- Preprint
(28180 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-3652', Reinhard Hinkelmann, 17 Jul 2026
-
RC2: 'Comment on egusphere-2026-3652', Anonymous Referee #2, 14 Oct 2026
General Comment
The manuscript presents a high-resolution rain-on-grid (RoG) simulation of the 2021 Ahr flood using the multi-GPU SERGHEI-SWE solver over approximately 900 km². The computational achievement is noteworthy, and the results demonstrate the potential of HPC-based hydrodynamic modelling for large-scale flood simulations. However, its central scientific claim is considerably broader than the evidence provided. I have substantial concerns regarding the scientific novelty and, more importantly, the extent to which the results support the central claim that hydrological models can be replaced. The ability to reproduce flood inundation using a standalone hydrodynamic model should not be confused with the ability to accurately represent the hydrological response of an entire catchment.
The results demonstrate that a standalone RoG configuration can reproduce important inundation characteristics of the 2021 Ahr flood with reasonable accuracy and within computationally feasible runtimes. They do not demonstrate that conventional hydrological models can generally be replaced, nor do they establish that the complete rainfall–runoff transformation is accurately represented across the catchment.
If additional analyses cannot be provided, the manuscript should be substantially reframed as a demonstration of computational feasibility and event-specific predictive performance, rather than as evidence that hydrological models can be replaced. In its present form, the manuscript provides convincing evidence of computational progress, but not yet of the broader hydrological modelling paradigm shift suggested by its title. The manuscript should be substantially reframed as a demonstration of computational feasibility and event-specific predictive performance, rather than as evidence that hydrological models can be replaced.
In its present form, the manuscript provides convincing evidence of computational progress, but not yet of the broader hydrological modelling paradigm shift suggested by its title.
Specific comments
1. Previous studies have already demonstrated that RoG approaches can effectively reproduce flash floods at catchment scale, providing a viable alternative to conventional hydrological–hydrodynamic coupling under specific conditions. The present study therefore does not introduce a fundamentally new modelling concept. Its main contribution lies in extending an existing modelling framework to a larger computational domain at metre-scale resolution, supported by GPU-based HPC. While technically valuable, this represents primarily a computational advancement rather than the conceptual breakthrough suggested by the title. The authors should more clearly distinguish their contribution from previous RoG applications and substantially reframe their paper.
2. The manuscript does not provide a direct comparison with a coupled hydrological–hydrodynamic modelling framework under equivalent conditions. Comparisons with previously published studies involve different modelling configurations, boundary conditions, and potentially different validation procedures. Moreover, some of the cited benchmarks are hydraulic simulations driven by prescribed hydrographs rather than complete coupled modelling chains. Consequently, the results demonstrate the feasibility of a standalone RoG simulation, but not its equivalence or superiority to conventional hydrological modelling. Either a consistent comparative experiment should be provided, or the central claim should be reformulated accordingly.
3. The adopted infiltration scheme assumes a temporally constant infiltration capacity and does not represent the evolution of soil moisture, subsurface flow, groundwater contributions, or interstorm processes. These limitations are particularly important because the hydrological response of a catchment depends on antecedent conditions and their evolution between successive rainfall events. Although the adopted simplifications may be acceptable for certain extreme floods, their applicability cannot be generalised to different rainfall regimes or continuous simulations. The replacement of hydrological models is not demonstrated. How would the proposed standalone framework represent the evolving hydrological state of the catchment between successive events without introducing additional hydrological process representations? This is a fundamental limitation of the claim that hydrological models can be replaced.
4. Although the computational domain covers approximately 900 km², the main validation datasets concern the lower Ahr Valley. Furthermore, flood-extent validation is restricted to a buffer around the main river, excluding much of the catchment where runoff generation occurs. Good agreement in downstream inundation extent does not necessarily demonstrate accurate representation of rainfall–runoff transformation, since compensating errors in runoff generation and routing may produce similar flood patterns. The reconstructed stage hydrograph also shows a 2–3 h anticipation of the peak, which deserves greater attention, particularly in view of the operational implications discussed by the authors. Additional validation against discharge hydrographs, preferably at multiple locations, would substantially strengthen the study. Otherwise, the conclusions should be restricted to the aspects of flood dynamics that have actually been validated.
5. The authors repeatedly emphasise that the simulations were performed without calibration. This is potentially an important aspect of the study, but I have some concerns on this. Several modelling choices appear to rely on previous investigations of the same flood event. For example, the adopted Manning parameterisation has been used in earlier applications, while the treatment of buildings is justified by reference to previous results indicating favourable performance for the Ahr Valley.
6. The computational results are among the strongest aspects of the manuscript. The ability to simulate a 96-hour event substantially faster than real time demonstrates the considerable potential of modern multi-GPU hydrodynamic solvers. However, computational speed alone does not establish operational forecasting readiness. A critical aspect, which I suggest discussing in detail, is that an operational forecasting system must also account for the availability and uncertainty of precipitation forecasts (as shown recently by several authors; see, for example, [1–3] mentioned at the end of my review), the specification of hydrological initial conditions, the evolution of catchment states between events, and the computational requirements associated with ensemble predictions. In the present study, the simulations are performed retrospectively using observed radar precipitation, so they do not constitute a complete operational forecasting framework. The authors should therefore distinguish between demonstrating computational feasibility within operationally relevant runtimes and demonstrating actual forecasting capability.
7. The conclusions are far from being generalizable. The study is based on a single extreme flood event in a single catchment. The Ahr catchment is characterised by steep topography, confined valleys, and a particularly severe rainfall event. These characteristics may favour an approach in which surface-flow dynamics dominate the simulated response. However, the relative importance of surface and subsurface processes varies considerably between catchments and events. The results presented here cannot establish whether the same modelling configuration would remain effective under different hydrological conditions, including moderate rainfall events, prolonged wet periods, or sequences of storms separated by interstorm intervals. This limitation is especially important given the general wording of the manuscript's title. The authors should avoid extrapolating the demonstrated performance to hydrological conditions that have not been investigated. A more defensible conclusion would be that high-resolution RoG hydrodynamic modelling represents a computationally feasible and potentially effective alternative for simulating certain classes of extreme rainfall-driven floods, provided that the dominant runoff-generation mechanisms can be adequately represented. But this is already known in the literature.
In conclusion, I appreciate the work carried out by the authors and congratulate them on the excellent results achieved in terms of computational speed. I believe this is a valuable contribution to the literature in this field; however, the core message presented in the title and propagated throughout the text cannot be accepted as it stands, as it is not fully supported by the findings presented in the study.
Recommendation: Major Revisions
Cited works
1. DOI: 10.1016/j.jhydrol.2026.135429
2. DOI: 10.1016/j.ijdrr.2026.105998
3. DOI: 10.1016/j.jhydrol.2024.132564
Citation: https://doi.org/10.5194/egusphere-2026-3652-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 239 | 129 | 30 | 398 | 28 | 23 |
- HTML: 239
- PDF: 129
- XML: 30
- Total: 398
- BibTeX: 28
- EndNote: 23
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The authors nicely demonstrate that hydrological models or combined hydrological-hydrodynamic models can be replaced by high-resolution (10 m and below) hydrodynamic models alone on High Performance Computers such as multi-GPUs. This is the case for flood investigations and for spatial scales of up to 1000 km² and time scales of several days, similar accuracies compared to the previous approaches are obtained and this approach has a couple of advantages, e.g. no model calibration or model coupling. Overall, this is a significant step forward for high-resolution 2D shallow water flood modelling. The paper is well written and structured to a large extent and only requires minor revisions.
Further minor points are in a pdf attached; no need to comment on them.