The manuscript presents an evaluation of a regional physics-based numerical weather prediction model (WRF) and three global data-driven weather models (AIFS, GraphCast and FuXi) over three high-impact flash-flood events in Luxembourg and the Greater Region. The study analyses the impact of data assimilation on WRF forecast skill, particularly for precipitation, and compares the different forecasting systems against surface observations using categorical verification scores for precipitation (based on a threshold of 1 mm per 6 hours) and continuous statistics for 2-metre temperature.

The overall objective of comparing the strengths and weaknesses of physics-based and AI-based weather forecasting approaches is timely and highly relevant for guiding future developments in numerical weather prediction. I consider the manuscript suitable for publication in Geoscientific Model Development after revisions addressing the comments below.

General comments:

1. In the Introduction (Line 41 and onwards), the three case studies are briefly described as flash-floods events in the area. However, throughout the manuscript, precipitation verification relies on categorial scores computed using a threshold of 1 mm per 6 hours. This threshold is appropriate for evaluating the spatial coherence of precipitation but is less informative regarding rainfall intensity. For example, the spatial comparison of the July 2021 event suggests that the AI models tend to smooth the precipitation field, whereas WRF (particularly with data assimilation) better captures the intensity of the event. Would the main conclusions still hold when considering higher precipitation thresholds that are more representative of extreme rainfall (e.g., 5, 10 or 20 mm per 6 hours)?
2. Regarding the comparison of forecast skill among the different models (WRF with and without DA, AIFS, GraphCast and Fuxi), I recommend including confidence intervals to support statements about differences in model performance. This is particularly important for Figures 4-7 and for the corresponding discussions in Sections 6.1.1, 6.1.2, 6.2.1, etc. In addition, regarding the verification metrics, please clarify if they are computed just in those days where the events occur (i.e., 21-23 July of 2016) or if they are computed in the entire forecasted cycle of 90 days. In the case of the latter, I would recommend analysing the metrics just during the event. This would help to focus on the modelling skill in the extreme events.
3. The manuscript clearly demonstrates the positive impact of data assimilation on WRF forecasts, particularly for precipitation. Have the authors assessed the sensitivity of the results to the data assimilation cycling frequency (e.g., 3-hourly or 1-hourly instead of 6-hourly)?
4. Could the authors clarify the independence of the verification dataset? It is not entirely clear whether all surface observations used for verification were also assimilated into WRF. This is particularly important when comparing WRF with AI models, since the latter ones do not ingest these observations, as they are initialised directly from ERA5. Indeed, comparing them with the WRF simulation without DA may provide a fairer benchmark.
5. The spatial comparison between radar observations and model forecasts for the July 2021 event (Fig. 9) is particularly informative. Likewise, the time series at the four selected stations (Fig. 10) reveal some differences between the radar estimates and the station measurements, especially at the Vatry station. Have the authors examined nearby stations to determine whether this discrepancy reflects a localised measurement issue or a limitation of the radar product? In addition, I suggest overlaying the observed stations precipitation measurements as coloured markers on the maps in Fig. 9 to facilitate a direct comparison between station observations, radar estimates and model forecasts.
6. Is the behaviour observed for the AI models during the July 2021 event (e.g., smoothing in AIFS and GraphCast, or the weak precipitation signal in FuXi) also found for the 2016 and 2018 events? Although radar data are unavailable for those cases, the comparison could still be performed using the station observations, as suggested in my previous comment.
7. The conclusions provide a clear summary of the main findings and briefly discuss the performance of the AI models relative to WRF. However, it should be acknowledged that the AI models evaluated here are global, whereas WRF is a regional dynamical model. This fundamental difference may partly explain the smoother precipitation fields produced by the AI models. Related to this, while the WRF simulations had a horizontal resolution of 12 km, the AI models had around 25 km. Will the results still hold if the WRF simulation were run at 25 km? Please, add some discussion around these points. It would also strengthen the discussion to compare these results with recent studies on regional or limited-area AI weather models specifically designed to the prediction of extreme events (e.g., Xu et al., 2025; Baño-Medina et al., 2025).

Specific comments:

Line 32: I suggest reframing this paragraph to motivate your analysis. In the first part, it is described one study that finds improvements from AI vs dynamical models for short-term regional forecasts, while latter is mentioned other which finds that extreme events are not so well captured by AI models. I would suggest using these different results to motivate your study, to try to clarify the strengths and weaknesses of each modelling framework by comparing them in three extreme events that took place in your research region.

Line 44: Is the choice of the WRF configuration, described as being tuned for this region, based on a previous study? If so, please provide the corresponding reference. In addition, how sensitive are the results to the selected physical parametrisations (Table 1), particularly the cumulus parametrisation?

Fig. 1: I suggest adding levels to the topography to better illustrate the reader the influence that the terrain can play in the study area.

Line 113: Why the WRF simulations use a 30-day window and not a shorter one? If there is a reason to this election, please justify it in the text.

Fig. 6: Why are the verification scores for WRF (with DA) different from those shown in Fig. 4?

Fig. 9: Could the authors include the WRF simulation without data assimilation in this figure? This would help to assess the impact of data assimilation on the spatial distribution and intensity of the predicted precipitation for this event.

Line 212: The discussion of the computational efficiency is very valuable. It would be helpful to include a quantitative comparison of the computational requirements (e.g., execution time) for the WRF and AI forecasts.

General comments

1 The comparison between WRF and AI models is interesting, but the experimental design may introduce an inherent imbalance in the initial conditions. The WRF experiments use ERA5 as boundary and initial conditions followed by a regional data assimilation system incorporating additional observations (SYNOP, TEMP, TAMDAR, GNSS ZTD), whereas the AI models are initialized directly from ERA5 analyses.

Although the authors discuss this issue in the conclusions, its impact deserves a more quantitative assessment. In particular, the comparison between observation-corrected WRF and AI models may partly reflect differences in the quality of the initial atmospheric state rather than differences between dynamical and AI forecasting approaches. I suggest including a more explicit comparison between WRF without DA and AI models, or quantifying the contribution of improved initial conditions relative to the forecasting model itself.

2 The AI models are re-initialised from ERA5 every 6 hours rather than being run in a fully autoregressive mode. This design avoids error accumulation but does not represent the operational use case of multi-step AI forecasting.

Please clarify why this evaluation strategy was selected and discuss how the results would change under free-running forecasts. Since one of the major advantages claimed for AI models is computational efficiency during long-range forecasting, it would be valuable to evaluate both single-step forecast skill and multi-step rollout degradation.

3 The WRF configuration uses a single 12 km domain. However, flash floods are often strongly influenced by convection-permitting processes and local topography. The manuscript should discuss whether the chosen WRF resolution is sufficient to represent convective rainfall mechanisms in this region. A comparison with higher-resolution convection-permitting WRF simulations would help determine whether the observed advantage of WRF for rainfall structure is related to physical modelling itself or simply to the spatial resolution difference.

4 The manuscript correctly mentions that AIFS benefits from ECMWF operational analyses. However, this issue deserves more emphasis because AIFS is not purely comparable with GraphCast and FuXi. AIFS uses fine-tuning with IFS analyses from 2016–2022, which overlaps with all three evaluation events. Therefore, the comparison does not only represent architecture differences but also differences in training data and operational assimilation systems.

5 Table 2 provides basic information about AI models, but important details for reproducibility are missing. Please consider providing

     exact model checkpoints/releases used;

     inference settings;

     precipitation variable definition;

     any preprocessing applied to ERA5 inputs;

     interpolation method from model grid to station locations.

6 Figure 9 provides qualitative comparison of rainfall fields. However, visual comparison alone is subjective. Please consider adding spatial verification metrics, such as Fractions Skill Score (FSS), object-based precipitation verification, spatial correlation...This would strengthen the conclusion that WRF better represents rainfall structures.