Toward AI-Driven Digital Twins for Metropolitan Floods: A Conditional Latent Dynamics Network Surrogate of the Shallow Water Equations

Technical Summary: Toward AI-Driven Digital Twins for Metropolitan Floods

1. Problem Statement

Urban and regional flooding is a rapidly intensifying hazard, yet high-fidelity hydrodynamic forecasting at the metropolitan scale remains computationally prohibitive for operational workflows requiring ensemble forecasting, scenario screening, or data assimilation. While GPU-accelerated solvers for the two-dimensional Shallow Water Equations (SWEs), such as SynxFlow, can produce deterministic forecasts, they require approximately 55 minutes to simulate a 96-hour event over a 4.2-million-active-cell metropolitan basin (the Des Plaines River basin at 30 m resolution). This latency scales linearly with the number of runs, making workflows like rainfall-forecast ensembles or climate-projection downscaling infeasible at native resolution.

Existing data-driven flood surrogates face three critical gaps when applied to operational metropolitan digital twins:

1. Scalability: Most architectures (GNNs, CNNs, VAEs, FNOs) have memory costs scaling with the full grid size, forcing coarsening or cropping of domains. They struggle with domains exceeding ~1 million cells.
2. Terrain-Aware Dynamics: Many models predict static inundation maxima or handle terrain only implicitly via convolutional biases, failing to explicitly capture the joint governance of time-varying rainfall and spatially heterogeneous static terrain (elevation, slope, roughness).
3. Observation Compatibility: Grid-based surrogates typically output raster data tied to the simulation mesh, requiring interpolation for sensor queries. They lack a compact, meshless latent state suitable for direct integration with sparse gauge observations for data assimilation.

2. Methodology: Conditional Latent Dynamics Network (CLDNet)

The authors propose CLDNet, a meshless neural surrogate that factorizes the spatio-temporal SWE solution into a low-dimensional latent dynamical system and a coordinate-based decoder.

• Latent Dynamics (Neural ODE): The model advances a low-dimensional latent state $s_t \in \mathbb{R}^{d_s}$ using a neural ordinary differential equation driven by rainfall forcing. The state evolution is defined by:
  $$s_{t_k} = s_{t_{k-1}} + \Delta t \cdot \mathcal{F}_{\theta_1}(s_{t_{k-1}}, r_{t_k})$$
  where $r_{t_k}$ is the rainfall forcing and $\mathcal{F}_{\theta_1}$ is a dynamics network (MLP).
• Terrain-Conditioned Decoder: Unlike standard Latent Dynamics Networks (LDNets) which rely solely on coordinates and the latent state, CLDNet introduces an explicit static terrain conditioning vector $\phi(\xi)$ concatenated with the Fourier-embedded spatial coordinate $\gamma(\xi)$ at every query point. The reconstruction network $\mathcal{R}_{\theta_2}$ decodes the conserved variables (water depth $h$, unit-width discharges $hu, hv$) as:
  $$\tilde{q}(t_k, \xi) = \mathcal{R}_{\theta_2}(s_{t_k}, [\gamma(\xi), \phi(\xi)])$$
  The conditioning vector $\phi(\xi)$ consists of standardized elevation, slope magnitude, and scaled Manning roughness. This allows the decoder to explicitly account for local hydraulic controls without memorizing the entire terrain field through the latent state.
• Meshless Training and Inference: Because the decoder operates pointwise, training and inference memory scale with the number of query points ($M$) rather than the full grid size ($|\Omega_h|$). The authors employ spatial subsampling, drawing a subset of active cells (e.g., $M=70,000$ for Illinois) per training step. This enables training on a single dual-GPU node for domains with >4 million active cells.
• Query Flexibility: The model can be queried at arbitrary coordinates, including exact USGS gauge locations $(lon, lat)$, without raster snapping or interpolation.

3. Datasets and Validation

The study evaluates CLDNet on two datasets:

1. Texas Benchmark: A synthetic, controlled dataset with a 250,000-cell domain (30 m resolution) and spatially uniform rainfall. Used to isolate the effect of terrain conditioning.
2. Des Plaines / Illinois Case Study: A realistic, metropolitan-scale dataset with 4,188,840 active cells (30 m resolution) driven by 114 real-rainfall storm forcings from the NCEP Stage IV archive (2002–2024).
   • Simulator Validation: The underlying SynxFlow simulator was validated against six USGS stream gauges during the April 2013 flood-of-record. After mean-recentering to remove datum offsets, the simulator achieved Nash–Sutcliffe Efficiency (NSE) values between 0.57 and 0.94, establishing it as a credible reference target for the surrogate.

4. Key Results

• Accuracy Gains:
  • Texas Benchmark: CLDNet reduced the aggregate test relative Root-Mean-Squared Error (rRMSE) from 31.09% (unconditional LDNet) to 14.44%. It outperformed VAE–ConvLSTM (32.04%) and FNO (18.10%) baselines.
  • Illinois Case Study: CLDNet roughly halved the error of the unconditional LDNet on held-out test trajectories (rRMSE reduced from 35.04% to 17.79%).
  • Flood Extent: At the 0.5 m inundation threshold, CLDNet achieved a Critical Success Index (CSI) of 86.04% on the Illinois dataset.
  • Gauge Performance: At USGS gauges, CLDNet improved the mean NSE from 0.523 to 0.711 and reduced peak-depth relative error by roughly half (15.85% to 9.67%).
• Computational Efficiency:
  • CLDNet produces a full 96-hour, 4.2-million-cell forecast in ~29 seconds.
  • This represents a ~115× speedup over the SynxFlow simulator (~55 minutes).
  • The speedup enables previously infeasible workflows, such as a 1,000-member rainfall ensemble (reducing runtime from ~38 days to ~8 hours on a single GPU).
• Scalability:
  • Unlike regular-grid baselines (VAE–ConvLSTM, FNO) which failed to train on the Illinois dataset due to memory constraints (requiring >1 TB of GPU memory), CLDNet trained successfully on a dual-GPU node by decoupling memory from grid size.
  • The model natively handles the irregular, non-rectangular geometry of the Des Plaines watershed without padding or masking.

5. Significance and Claims

The paper positions CLDNet not as a complete operational digital twin, but as a fast hydrodynamic surrogate layer essential for building AI-driven flood digital twins. Its primary contributions are:

1. Metropolitan Scalability: Demonstrating that neural surrogates can scale to >4 million active cells on commodity hardware via meshless, pointwise decoding.
2. Terrain Conditioning: Proving that explicit conditioning on static terrain features (elevation, slope, roughness) is critical for regional-scale accuracy, significantly outperforming unconditional latent dynamics.
3. Observation-Ready Interface: Providing a compact latent state and a meshless query operator that allows direct evaluation at exact gauge coordinates, structurally enabling future integration with data assimilation algorithms (e.g., ensemble score filters).
4. Operational Viability: Establishing that the ~115× speedup makes ensemble forecasting and scenario screening tractable at native metropolitan resolution, a prerequisite for real-time, observation-coupled flood management.

The authors note limitations, including the current lack of online observation assimilation (positioned as a future extension) and the fact that discharge predictions ($hu, hv$) remain less accurate than water depth predictions, particularly near wet/dry interfaces. However, the work establishes a foundational architecture for the next generation of operational flood forecasting systems.