Research map / Probabilistic and scientific modeling / Differential equations and PDEs
Pluvial Flood Emulation with Hydraulics-informed Message Passing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ComGNN, a hydraulics-informed GNN on a D8 flow-direction graph that predicts pluvial flood water depths auto-regressively via two-stage water retention and dispersion message passing inspired by shallow-water equations.
Problem
Simulating spatio-temporal pluvial flood water depths over complex real-world topography under sparse, time-varying precipitation, while incorporating hydraulics physics into data-driven models.
Approach
A two-stage graph message passing on a static D8 flow-direction graph: rain is first retained per cell (bucket), then water is dispersed via momentum-inspired flow messages gated by sigmoid of water level and mass-conservation depth updates, trained with a combined L1/L2 loss plus a non-negativity penalty.
Key results
- ComGNN achieves RMSE 0.1328 at t=20 vs 0.1571 for the best non-ablated competitor ComGNN- and 0.1597 for MeshGraphNet, and NSE 0.8866 at t=20 (Table 1).
- On unseen regions and precipitations (Table 4), ComGNN attains RMSE 0.0809 at t=20 versus 0.0931 for MeshGraphNet.
- On a flat surface (Table 3), ComGNN reaches NSE 0.9231 versus 0.8697 for MeshGraphNet, while the baselines degrade on real irregular topography.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a water-retention stage before water propagation (ablation) | ComGNN · All test regions and precipitation events | RMSE: .1571 -> .1328 (t=20); .4830 -> .3615 (t=40) | Helped |
| Use hydraulics-informed water-dispersion message passing instead of a graph attention network (ablation) | ComGNN · All test regions and precipitation events | RMSE: .2155 -> .1328 (t=20); .6103 -> .3615 (t=40) | Helped |
| Use hydraulics-informed water-dispersion message passing instead of a graph convolution network (ablation) | ComGNN · All test regions and precipitation events | RMSE: .1874 -> .1328 (t=20); .5482 -> .3615 (t=40) | Helped |
| Initialize simulation with true water depths at t=10 rather than t=0 | Flood emulation baselines · All test regions and precipitation events | RMSE: Table 1 and Table 2 report different initial-condition settings; no paired change value is stated | Mixed |
| Use a flat surface rather than real-world irregular topography | MeshGraphNet, MP-PDE, and FNO · Flat surface | NSE: +27% in NSE for MeshGraphNet; +14% in NSE for MP-PDE; +20% in NSE for FNO | Helped |
Surprising findings
- Rain-Incr, a one-parameter baseline, is competitive with sophisticated methods such as MP-PDE and GCN.
- Convolution-based methods (ConvLSTM, U-net, FNO) improve markedly when initialized at t=10 while GNN-based methods, including ComGNN, improve less, attributed to the D8 graph representation.
Limitations
- Fixed D8 flow-direction graph limits performance when water propagates in many directions, especially from later initial states.
- Evaluated only on Harris County, TX regions with LISFLOOD-FP synthetic labels rather than observations.
- Simulation lead time limited to 40 steps by GPU memory.
Experimental setup
Datasets: LISFLOOD-FP simulations over 9 Harris County, TX sub-watersheds with 7 historical rainfall events from MRMS-GC QPE
Benchmarks: LISFLOOD-FP-simulated Harris County flood dataset
Metrics: RMSE NSE Pearson r CSI
Baselines: U-net ConvLSTM MP-PDE FNO MeshGraphNet MLP GCN GAT Rain-Incr
Builds on: MeshGraphNet MP-PDE Method of lines LISFLOOD-FP
Closest papers by meaning
- Physics-Informed Graph Neural Networks for Water Distribution Systems (AAAI 2024 · similarity 0.87)
- FIDLAR: Forecast-Informed Deep Learning Architecture for Flood Mitigation (AAAI 2025 · similarity 0.87)
- RiverMamba: A State Space Model for Global River Discharge and Flood Forecasting (NeurIPS 2025 · similarity 0.86)
- The Merit of River Network Topology for Neural Flood Forecasting (ICML 2024 · similarity 0.85)
- Physics-Guided Fair Graph Sampling for Water Temperature Prediction in River Networks (AAAI 2025 · similarity 0.84)
- PINP: Physics-Informed Neural Predictor with latent estimation of fluid flows (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.