| 09 Oct 2025
High-performance coupled surface-subsurface flow simulation with SERGHEI-SWE-RE
Abstract. This work presents SERGHEI-SWE-RE, a performance-portable, parallel model that couples a fully dynamic two-dimensional Shallow Water Equation (SWE) solver with a three-dimensional Richards Equation (RE) solver within the Kokkos framework to simulate surface–subsurface flow exchange. The model features a modular architecture with sequential coupling strategy, supporting both synchronous and asynchronous executions of surface and subsurface modules. The SERGHEI-SWE-RE model is validated against five benchmark problems incorporating stationary and fluctuating free-surface tests, a tilted v-catchment, a lateral-flow slope without ponding, and a heterogeneous superslab. The results demonstrate good agreement with established models. Asynchronous coupling reduces wall-clock time by up to about 60 % in the superslab case while preserving simulation accuracy. Strong and weak scaling tests on multiple Intel Xeon CPUs and NVIDIA GPUs reveal robust portability, with near-ideal RE scaling and less-satisfactory SWE scaling at high GPU counts, suggesting future improvements on differentiated meshes or more advanced domain decomposition strategies. Overall, the results presented establish SERGHEI-SWE-RE as an efficient, flexible and scalable model for integrated surface-subsurface flow simulations.
This manuscript presents SERGHEI-SWE-RE, a coupled surface–subsurface model that links a fully dynamic 2D shallow-water solver with a 3D Richards solver. Overall, the work is solid and clearly within the scope of GMD. The model is well described, the benchmark cases are appropriate, and the scaling analysis across several HPC systems is particularly useful. The asynchronous coupling strategy is also a strength of the paper, and the manuscript is generally clear and well organized. I find the manuscript suitable for publication after a set of minor revisions. My comments below are intended to improve clarity and strengthen the presentation.
Major Comments:
The sequential coupling strategy is well explained, but it would be helpful to briefly discuss its potential limitations. In particular, readers would benefit from understanding situations in which asynchronous coupling may introduce small inaccuracies, or cases where a fully coupled approach might be more appropriate. A short clarification in Section 2.2 or in the Discussion would be sufficient. For example, in Case 5 the authors test two maximum RE time steps, which affects the overall runtime almost linearly. While I am not requesting additional experiments, it would strengthen the paper to comment on how sensitive the results are to the chosen coupling time window. How does varying the subsurface time step influence accuracy and computational cost? A brief discussion of this trade-off would help contextualize the asynchronous approach.
One overall question the scalability discussion does not fully address is the performance impact introduced by the coupling itself. In other words, how much does the integration of the SWE and RE solvers degrade performance compared to running the two components independently? A brief quantitative or qualitative assessment of this overhead, whether due to synchronization, data exchange, or load imbalance, would help readers better understand the true cost of coupling and the efficiency of the current implementation.
Rainfall is always applied to the SWE, even when no ponding exists. Since this is a modeling choice and may not hold in all situations, it would be good to clarify the limitations of this assumption. The Case 4 comparison shows the approach works well there, but a sentence noting scenarios where this may be less appropriate would be helpful.
Specific and Minor Comments:
Line 125, “that” to “than”
Figure 3 and the accompanying explanation are hard to follow. In the figure, if the intervals are meant to represent the time steps Δt, their lengths should be consistent. It is also not clear what the shaded/boxed “tsub” period represents. Does this interval mark the window during which the SWE and RE solvers exchange information? If so, this should be stated more explicitly in both the figure and the text.
In Section 2.3, the description of domain decomposition suggests that the model uses a structured grid without regional refinement. If this is indeed the case, it would be helpful to state this explicitly. Clarifying this will make it easier for readers to understand the limitations in the current domain decomposition strategy that you discuss later in the manuscript.
Line 174-176. The author should either report the results even in the supplement, or they should not use such a statement to support their conclusion.
Line 179, what is the range of y for the inlet boundary?
Line 246 “Fig. 14(b)” -> “Fig. 15(b)”
Lines 247–248: The explanation provided is reasonable, but I suggest elaborating a bit more here. Since simplified governing equations such as the kinematic and diffusive wave formulations are widely used in rainfall–runoff modeling, the distinction you highlight is important. This result that full SWE produces systematically larger ponding depths compared to the simplified approaches deserves to be emphasized more clearly, as it is likely to be of broad interest to the hydrologic modeling community.
For Figures 15 and 19, several curves overlap almost exactly, which makes them hard to distinguish. In Fig. 15, the blue and red lines lie directly on top of each other; using different line thicknesses (or slightly different styles) would make the overlap clearer. Similarly, in Fig. 19, the black circular markers are difficult to see because they coincide with another line. Adjusting line weights or marker sizes would improve readability in both figures.