Tony Kim
February 13, 2025 21:37
BRLI and TOULOUSE INP uses NVIDIA Modulus to create AI -based flood models to greatly improve real -time flood prediction and risk relief.
The floods are a serious threat to 1.5 billion people around the world, resulting in an economic damage of up to $ 25 billion every year. According to NVIDIA’s official blog, we have developed AI -based solutions that use BRLI and NVIDIA modulus according to urgent need for efficient flood prediction. This innovative approach promises to revolutionize the real -time flood forecast by dramatically reducing the calculation time.
Challenge of traditional flood prediction
Traditional flood predictions rely on physical -based numerical simulations that are computational and takes a lot of time. This method can take several hours to simulate flood events and limit utilities in real -time scenarios. This bottleneck has prevented the development of a reactive flood warning system that can provide timely executable insights during the ongoing event.
AI -based solution
To overcome these limitations, the BRLI and TOULOUSE INP designed an AI system that replaces traditional physical solvifications through Aniti Research Institute. By using NVIDIA MODULUS on the Earth-2 platform, the team trained the AI model to imitate the solver to enable the rapid evaluation of the flood scenario.
The AI model, which has been trained for the detailed physical models provided by BRLI, can imitate a few hours of floods in a single GPU in a few seconds. This innovation greatly improves the potential of real -time predictions and decision -making in areas where floods are easy to occur.
Implementation and testing
The AI -based system focuses on Têt River Basin in southern France, using detailed meshes that include important terrain and engineering functions. This system uses the NVIDIA modulus to train models for custom data to optimize the complex space and temporal mechanics that are important for accurate flood predictions.
Training for the NVIDIA A100 Tensor Core GPU has been carried out, increasing the speed and up to 30 minutes of predictions up to several hours ahead. The accuracy of the model has been verified and reliable using a metric such as an average square error (MSE) and CSI (CSI).
Influence and future prospects
The resulting surrogate GNN model can perform a 6 -hour prediction within only 19 milliseconds in a single NVIDIA A100 GPU, contrasting with the 12 -hour CPU time required for traditional methods. This efficiency enables real -time flood modeling without damaging the complexity of the simulation.
This development not only shows the functions of NVIDIA Modulus in setting and training the AI architecture, but also sets precedents for similar applications in various industries. The success of this project opens up the way to improve the ability to respond efficiently to natural disasters by integrating the AI model into operating disaster relief services.
As the BRLI and TOULOUSE INP improves the model, integrating AI into engineering toolchains will be more and more executable. This development refers to the main stage of flood risk management and provides expandable and efficient solutions for continuous global challenges.
For more information, visit the NVIDIA blog.
Image Source: Shutter Stock