Pluvial Flood Emulation with Hydraulics-informed Message Passing
Authors: Arnold Kazadi, James Doss-Gollin, Arlei Lopes Da Silva
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results on a dataset covering 9 regions and 7 historical precipitation events demonstrate that our model outperforms the best baseline, and can capture the propagation of water flow more effectively, especially at the very early stage of the flooding event when the amount of water in the domain is scarce.Extensive experiments show that these features enable our method to simulate flooding events much more accurately than existing approaches, including modern ML models for physics-based simulation. |
| Researcher Affiliation | Academia | 1Department of Computer Science, Rice University, Houston, TX, USA 2Department of Civil and Environmental Engineering, Rice University, Houston, TX, USA. Correspondence to: Arnold Kazadi <akn7@rice.edu>. |
| Pseudocode | No | The paper describes the model using equations and textual descriptions but does not provide pseudocode or a clearly labeled algorithm block. |
| Open Source Code | Yes | The codebase and datasets used in our experiments can be accessed via the repository at https://github.com/kanz76/Com GNN.git |
| Open Datasets | Yes | The codebase and datasets used in our experiments can be accessed via the repository at https://github.com/kanz76/Com GNN.gitExperiments are based on the simulations from the hydrodynamic model LISFLOOD-FP (Shaw et al., 2021).We consider 9 sub-watershed regions from Harris County, in Texas (see Figures 4 and 5 and Table 6 in Appendix A.3). For each of these regions, simulations were run using 7 historical rainfall events (based on the flood history in Harris County1 collected from NOAA NEXRAD radar precipitation records from the Multi-Radar Multi-Sensor Gauge Corrected (MRMS-GC) Quantitative Precipitation Estimation (QPE) product (Martinaitis et al., 2020). |
| Dataset Splits | Yes | There are 63 combinations coming from the 9 sub-watershed regions and 7 rainfall events, of which 9 combinations were used for the training, 3 combinations were used for validation, and the remainder were used for testing. |
| Hardware Specification | Yes | The simulation lead time was set to 40, the largest we could train on a single NVIDIA GPU Ampere A40. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies used in implementing the proposed model (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | The learning rate was set to 1e-4 for all the models.Com GNN showed better performance with a 3-layer MLP for Eq. 3, one layer of Eq. 8, and 2 layers of Eq. 10). tanh was used as the activation function and all the layers were implemented with 32 neurons. |