Ghomash et al. (2026) Enabling real-time high-resolution flood forecasting for the entire state of Berlin through multi-GPU accelerated physics-based modeling

Identification

• Journal: Natural hazards and earth system sciences
• Year: 2026
• Date: 2026-01-13
• Authors: Shahin Khosh Bin Ghomash, Siqi Deng, Heiko Apel
• DOI: 10.5194/nhess-26-85-2026

Research Groups

• Section Hydrology, GFZ German Research Centre for Geosciences, Potsdam, Germany
• Department of Earth System Science, Stanford University, Stanford, USA
• Institute of Geo-Hydroinformatics, Hamburg University of Technology, Hamburg, Germany

Short Summary

This study demonstrates the operational feasibility of real-time, high-resolution pluvial flood forecasting for large urban areas like Berlin using the multi-GPU accelerated hydrodynamic model RIM2D. It shows that RIM2D can deliver detailed flood simulations and impact estimates significantly faster than real-time, making it suitable for early warning systems and comprehensive urban flood risk management.

Objective

• To investigate whether high-performance computing (HPC)-enabled shallow water solvers, specifically RIM2D with multi-GPU acceleration, can achieve sufficient accuracy and lead time to support early flood warning systems over large urban domains such as the entire state of Berlin.
• To evaluate if such models can be run with sufficiently short computation times to support real-time, impact-based flood forecasting and if multi-GPU configurations are essential for this.

Study Configuration

• Spatial Scale: Entire state of Berlin (891.8 km²), with simulations at 2 meter, 5 meter, and 10 meter spatial resolutions.
• Temporal Scale: 48-hour simulation for the June 2017 pluvial flood event; 1-hour simulation for a standardized 100-year return period (HQ100) rainfall event.

Methodology and Data

• Models used: RIM2D (Rapid Inundation Model 2D), a multi-GPU accelerated 2D hydrodynamic model written in CUDA FORTRAN, which solves the local inertia form of the Shallow Water Equations.
• Data sources:
  • Topography: DGM1 digital elevation model (DEM) of Berlin (Geoportal Berlin/ATKIS®DGM), resampled to 2 m, 5 m, and 10 m resolutions.
  • Buildings: Building outlines from OpenStreetMap, treated as closed reflective boundaries.
  • Roughness: Manning roughness values assigned based on 2020 land cover classification of Germany (derived from Sentinel-2 data).
  • Infiltration: Soil type from Umweltatlas Berlin/Bodenkundliche Kennwerte 2020 combined with Imperviousness Density percentage raster (Copernicus Land Monitoring Service, 10 m resolution).
  • Sewer Drainage: Capacity-based approach, estimated from KOSTRA database (DWD, 2017) for a 2-year, 15-minute design rainfall and DWA-A 118 guide, resulting in a mean capacity of 21 mm h⁻¹.
  • Precipitation Input: RADOLAN radar data (German Weather Service DWD) for the June 2017 event (1000 m × 1000 m spatial, hourly temporal); uniform rainfall intensity of 100 mm h⁻¹ for 1 hour for the HQ100 event.
  • Validation Data: Volunteered Geographic Information (VGI) including photos and videos from residents for the June 2017 event (19 locations); Berlin city’s official flood hazard maps for selected catchments (Obersee, Flughafensee, Niederschönhausen Ost) based on a 100-year return period rainfall event.
  • Impact Assessment: 2020 WorldPop population density dataset for Germany (100 m resolution).

Main Results

• For the 48-hour June 2017 flood event using 8 NVIDIA A100 GPUs, RIM2D simulation runtimes were: 8 minutes at 10 meter resolution, 34 minutes at 5 meter resolution, and approximately 5.5 hours at 2 meter resolution.
• Multi-GPU processing is essential for high-resolution simulations (e.g., 2 meter or finer) and for making simulations at resolutions finer than 5 meter computationally feasible for operational flood forecasting.
• Runtime improvements become marginal beyond 4 GPUs for 5 meter and 10 meter resolutions, and beyond 6 GPUs for 2 meter resolution, indicating an optimal balance between computational nodes and raster cells.
• RIM2D achieves significantly faster-than-real-time performance: 347 times faster at 10 meter resolution (8 GPUs), 84 times faster at 5 meter resolution (8 GPUs), and 8.7 times faster at 2 meter resolution (8 GPUs).
• Validation against VGI for the June 2017 event showed strong spatial alignment of simulated flood extents with observations, and generally small underestimation of water depths, within an acceptable range for uncalibrated models.
• Comparison with official city flood hazard maps for the HQ100 event demonstrated strong agreement in flood extent and water depths across various performance metrics (CSI, HR, FA, EB, BPI, RMSE, Bias), with minor resolution-dependent improvements.
• RIM2D can generate high-resolution (2 meter) flood hazard maps for the entire Berlin area in under 30 minutes for the HQ100 scenario, a significant advancement over current practices limited to smaller catchments.
• Simulations for the June 2017 event at 5 meter resolution showed most areas experienced water depths between 5 centimeters and 20 centimeters, with localized hot-spots reaching approximately 50 centimeters, and generally low flow velocities.
• Impact assessment revealed a higher concentration of affected individuals (water depth > 0.1 meter) in the city center due to higher population density.

Contributions

• Demonstrates the technical feasibility of real-time, high-resolution, state-scale pluvial flood forecasting for large urban areas using multi-GPU accelerated physics-based hydrodynamic models.
• Highlights the critical role of multi-GPU computing in achieving operationally relevant runtimes for high-resolution simulations (finer than 5 meters) across extensive urban domains, overcoming single-GPU memory and performance limitations.
• Provides a robust framework for generating spatially explicit impact indicators (e.g., maximum water depth, flow velocity, time of peak inundation, affected population) that enhance disaster management and provide more actionable information than traditional coarse-resolution alerts.
• Enables the integration of detailed, real-time flood forecasting into urban flood risk management, supporting faster, data-driven decision-making and more effective risk mitigation.
• Offers a cost-efficient and accessible solution for flood forecast centers, leveraging readily available GPU hardware and cloud-based HPC services.

Funding

• EU Horizon 2020 (grant no. 101073978 HORIZON-CL3-2021-DRS-01)
• GFZ Helmholtz Centre for Geosciences (article processing charges)
• Ministry of Science, Research and Culture of the State of Brandenburg (MWFK) (for high-performance computing resources)

Citation

@article{Ghomash2026Enabling,
  author = {Ghomash, Shahin Khosh Bin and Deng, Siqi and Apel, Heiko},
  title = {Enabling real-time high-resolution flood forecasting for the entire state of Berlin through multi-GPU accelerated physics-based modeling},
  journal = {Natural hazards and earth system sciences},
  year = {2026},
  doi = {10.5194/nhess-26-85-2026},
  url = {https://doi.org/10.5194/nhess-26-85-2026}
}