Flooding poses a significant threat to 1.5 billion people, making it the most common cause of major natural disasters. Floods cause up to $25 billion in global economic damage every year. Flood forecasting is a critical tool in disaster preparedness and risk mitigation. Numerical methods have long been developed that provide accurate simulations of river basins. With these, engineers such as those at the consulting firm BRLi study different rainfall scenarios, and provide valuable assessments of flood risks, mitigation strategies, and disaster relief plans to local stakeholders. Accurate predictions can save lives, protect infrastructure, and reduce economic losses.Yet these methods use physics-based numerical simulations that are computationally intensive. This approach often takes many hours to simulate a flooding event, even on many parallel processors. It’s not suitable for real-time forecasting of ongoing events. This limitation has long been a bottleneck in providing rapid, actionable flood warnings tailored to a given event, and has prevented the development of real-time forecasting systems.To address this issue, a joint team at BRLi and National Polytechnic Institute of Toulouse (Toulouse INP), through a collaboration in the ANITI research institute, has designed an end-to-end AI approach that fully replaces the physics-based solver. This team includes expertise in both hydraulic modeling and AI methods for physics, enabling this interdisciplinary work. Hydrology experts from BRLi provided detailed physics models for a given basin to produce a database of floods. The Toulouse INP team then used these to train an end-to-end AI system to emulate the solver leveraging NVIDIA Modulus from the Earth-2 platform. Earth-2 is a digital twin cloud platform to develop AI-augmented forecasting and visualization pipelines for weather and climate applications. The team used the optimized training pipelines and parallel training capabilities along with the built-in deep learning models provided by Modulus. The resulting solver successfully emulates several hours of flooding ahead in mere seconds on a single GPU, opening the way for real-time forecasting. This is a groundbreaking approach leveraging AI to revolutionize flood forecasting, dramatically reducing computation time. AI-based flood forecasting system The joint team focused on the Têt River basin, in the south of France. Significant engineering efforts have led to detailed unstructured meshes of this region that encompass many important topographic and engineering features like bridges, dikes, and water retention basins. Detailed numerical simulations using the open-source Telemac-Mascaret code were run with a variety of water input profiles to produce a database of flooding events.For training, the INP team used NVIDIA Modulus as part of the Earth-2 platform, which provides a suite of tools for enabling AI powered solutions in the climate and weather domain. The team used one of the built-in model architectures in Modulus for fast and effortless training with their custom data. To explore what AI models can enable on weather and climate forecasting, check out the FourCastNet and CorrDiff NVIDIA NIM microservices in a browser-based experience. The INP team chose one of the graph neural network (GNN) architectures that has shown impressive performance in describing atmospheric dynamics and adapted it to the Têt River basin. This approach enables the model to capture complex spatial relationships and temporal dynamics crucial for accurate flood prediction. The AI model was trained to emulate the intricate hydrological processes that lead to flooding, by imitating the data from the detailed numerical simulations. Training was done on up to 14 NVIDIA A100 Tensor Core GPUs in parallel on the CALMIP Turpan supercomputer, achieving a near-linear speedup with a scaling efficiency of up to 91%. Using an autoregressive approach, the model can predict flooded areas and flow rates in 30-minute increments up to several hours into the future.To address the challenge of capturing long-range dependencies without substantially increasing the number of message-passing layers, the team created progressively coarser representations of the original mesh. These coarser meshes were connected to the fine mesh using a KD-Tree, forming a multimesh. This strategy expanded the model’s receptive field, enhancing its ability to simulate large-scale flood dynamics while maintaining computational efficiency.During training, the model optimization was guided by the mean squared error (MSE) loss function to ensure accurate predictions. For testing, the evaluation metrics included the L1 error to measure the average prediction deviation and the critical success index (CSI), which was computed at a 5 cm threshold. The CSI quantified the model’s ability to correctly detect and classify flooded areas by comparing the intersection of predicted and observed flooded cells to their union.Surrogate GNN modelThe final model can perform a 6-hour prediction in 19 ms on a single NVIDIA A100 80 GB GPU. In comparison, a similar simulation with the physics-based numerical model requires 12 hours of CPU time and at least 25 minutes of time-to-solution on 28 CPUs. What’s more, the method leverages the true mesh developed by the hydraulic engineering team with all its complexity. This is a true breakthrough for real-time flood modeling, for which AI attempts often require simplified configurations and meshes.This achievement also showcases how NVIDIA Modulus enabled the setup and training of cutting-edge AI architectures directly on engineering meshes. This example can be replicated for many problems across multiple industries, simply by providing simulation data.Figure 1. The surrogate GNN model can perform a 6-hour flooding prediction in 19 ms on a single NVIDIA A100 80 GB GPUConclusionBuilding on these exciting results, the team is currently refining the training data and evaluation metrics to ensure that the level of quality produced by the model matches the physics-based approach, including in rare events. Once fully validated, the model will be considered for integration into the engineering toolchains at BRLi. As a first step, it will be used to produce large ensembles of thousands of runs to obtain uncertainty quantification in basin studies, significantly improving upon the state of the art which relies on only a handful of simulations. Next, working with operational disaster relief services will be crucial to find the optimal data sources to ingest and feed to a real-time forecasting system, and how the data should be sent back to them.Earth-2 is an open platform and NVIDIA Modulus is an open-source project to support the growing physics AI community. To learn how NVIDIA Modulus can help your physics AI research projects, explore the NVIDIA/modulus GitHub repo.
