Flood Modelling: Operational Bottleneck and How to Address It

In our earlier discussion on Flood Early Warning System (FEWS) from the Global and Indonesia Perspective, there is a significant operational bottleneck due to the substantial computational resources required on conventional flood modelling with Physical-Based Numerical Models (PBNM).

These models, while offering detailed simulations, often necessitate extended processing times—sometimes several hours—which hampers the system’s ability to provide timely and continuously updated forecasts, especially during rapidly evolving flood events.

Building upon that foundation, this article delves into strategies to mitigate the computational demands of PBNM in flood modelling. Specifically, We will explore various techniques and assess their applicability within the Indonesian landscape, aiming to enhance the efficiency and responsiveness of flood forecasting systems across the archipelago.

I. Massive Parallelization: Running Flood Modelling on Multiple Cores

A. What Is

Think of flood modelling like calculating the total rainfall across 1,000 villages. Doing it yourself would take a full day. But if you had 100 friends, and each calculated rainfall for 10 villages, you’d be done in minutes. This is the principle behind massive parallelization. As shown in Figure 1, a large task is divided into smaller tasks, which are then processed simultaneously using many CPUs or GPUs—either on a High-Performance Computing (HPC) cluster or cloud-based infrastructure.

In flood modelling, the domain (e.g. a river basin or city) is split into grid cells, and the physical equations for water flow are solved in parallel. The total simulation time can be reduced drastically—from hours to minutes—without changing the underlying physical model¹.

B. Advantages & Disadvantages

C. Global Application

The use of massive parallelization has been widely adopted in flood modelling around the world. For instance, Liu² and Chen³ developed a parallelized flood forecasting and warning platform based on high-performance computing (HPC) clusters, which significantly improved forecasting capabilities for flash floods modelling across China.

Similarly, BBWS Citarum also has applied GPU techniques to enhance the efficiency of pluvial flood modelling in urban areas. These studies give a vivid glimpse on how massive parallelization can greatly improve computational performance in real-world flood forecasting applications

II. Adaptive Mesh Refinement: Improving Simulation Detail Where It Matters Most

A. What Is

In flood modelling, we divide the study area into a grid (or mesh), and simulate water movement cell by cell. A finer mesh (smaller grid cells) gives more detailed results, especially in complex urban areas or along riverbanks—but also increases computational time significantly. A coarser mesh is faster, but less accurate.

Adaptive Mesh Refinement (AMR) as seen in Figure 2, is a smart approach that dynamically adjusts the mesh resolution during the simulation. The model automatically increases resolution where more detail is needed—such as around floodplains, steep slopes, or built-up areas—and uses coarser resolution elsewhere. This way, you save computational power without sacrificing accuracy where it matters most.

Think of it like a camera that automatically zooms in on areas with a lot of movement, and zooms out where things are calm.

B. Advantages & Disadvantages

C. Global Application

Compared to massive parallelization techniques, Adaptive Mesh Refinement (AMR) requires more specialized model configurations and a deeper understanding of dynamic grid behavior, making it somewhat more complex to implement.

Despite this added complexity, numerous studies have demonstrated the effectiveness of AMR in accelerating flood simulations while maintaining high levels of accuracy. For example, Gong⁴ applied AMR in urban flood modelling in Japan, achieving efficient computation with detailed spatial resolution. Similarly, Hu⁵ utilized this method in the United Kingdom to improve simulation performance in dense urban environments.

III. Surrogate Modelling: Mimicking Flood Modelling with Machine Learning

A. What Is

Surrogate modelling is a technique that creates a simplified, data-driven approximation of a complex physical model. Instead of solving detailed physical equations each time (as in traditional flood models), a surrogate model learns the input–output relationships from previous simulations or real data and uses that learning to make predictions much faster.

These models are typically built using Machine Learning (ML) techniques such as Artificial Neural Networks (ANNs), Support Vector Machines (SVMs), or Gaussian Processes. Once trained, a surrogate model can predict flood extents, water levels, or inundation timing in seconds or minutes, making them especially useful for real-time applications.

Think of it like a flight simulator—it doesn’t calculate every force acting on a plane in real time, but it gives a realistic output based on prior knowledge of how planes behave.

B. Advantages & Disadvantages

C. Global Application

Surrogate models have gained traction as a complement or replacement for traditional hydrodynamic models, especially in early warning, uncertainty analysis, and real-time flood forecasting. A notable example is the study by Quang Thanh Dang⁶, which evaluated various machine learning algorithms using a synthetic dataset representing both pluvial and coastal flood scenarios.

The results demonstrated that Neural Networks, Gaussian Processes, and Random Forests outperformed other models in terms of predictive accuracy and computational efficiency, highlighting their potential for fast and reliable flood prediction.

IV. Conclusion

What can we take away from this? As Indonesia continues to face increasingly frequent and severe flood events, the demand for faster, more reliable flood modelling has become critical. While conventional physics-based numerical models provide high accuracy and physical realism, their intensive computational requirements often limit their use in real-time forecasting—precisely when speed is most essential.

This article has presented three promising solutions to address this challenge: Massive Parallelization, Adaptive Mesh Refinement (AMR), and Surrogate Modelling. Each technique offers unique strengths and trade-offs, and their effectiveness depends on factors such as available infrastructure, data quality, and specific modelling objectives.

For Indonesia, leveraging these methods—individually or in combination—holds significant potential to enhance the performance of flood forecasting systems, making them more responsive, scalable, and operationally effective. Embracing these innovations is not just a technical improvement, but a necessary step toward building flood-resilient communities and climate-adaptive infrastructure across the archipelago.

References

1. P. Luo et al., ‘Urban flood numerical simulation: Research, methods and future perspectives’, Environmental Modelling & Software, vol. 156, p. 105478, Oct. 2022, doi: 10.1016/j.envsoft.2022.105478. ↩︎
2. T. Chen et al., ‘High-performance computing in urban flood modeling: A study on spatial partitioning techniques and parallel performance’, Journal of Hydrology, vol. 649, p. 132474, Mar. 2025, doi: 10.1016/j.jhydrol.2024.132474. ↩︎
3. R. Liu et al., ‘A Parallel Flood Forecasting and Warning Platform Based on HPC Clusters’, presented at the HIC 2018. 13th International Conference on Hydroinformatics, pp. 1232–1223. doi: 10.29007/5vfl. ↩︎
4. W. Gong, Y. Shimizu, and T. Iwasaki, ‘A Case Study of Flood Modeling with Adaptive Mesh Refinement’. ↩︎
5. R. Hu, F. Fang, P. Salinas, and C. C. Pain, ‘Unstructured mesh adaptivity for urban flooding modelling’, Journal of Hydrology, vol. 560, pp. 354–363, May 2018, doi: 10.1016/j.jhydrol.2018.02.078. ↩︎
6. T. Q. Dang et al., ‘Application of machine learning-based surrogate models for urban flood depth modeling in Ho Chi Minh City, Vietnam’, Applied Soft Computing, vol. 150, p. 111031, Jan. 2024, doi: 10.1016/j.asoc.2023.111031. ↩︎