Toward AI-Driven Digital Twins for Metropolitan Floods: A Conditional Latent Dynamics Network Surrogate of the Shallow Water Equations
Toward AI-Driven Digital Twins for Metropolitan Floods: A Conditional Latent Dynamics Network Surrogate of the Shallow Water Equations
Authors: Phillip Si, Yuan Qiu, Omar Sallam, Jeremy Feinstein, Ziang He, Eugene Yan, Peng Chen Date: 2026-05-13 Paper ID: openalex:2605.13761
Summary
This paper introduces the Conditional Latent Dynamics Network (CLDNet) to accelerate metropolitan flood forecasting by acting as a surrogate for the shallow water equations. CLDNet utilizes a latent neural ODE to capture rainfall-driven dynamics and a coordinate-based decoder conditioned on terrain features to reconstruct hydraulic states without reliance on fixed grids. Evaluated on the Des Plaines River basin, the model achieves a 115x speedup over numerical solvers while providing superior accuracy and flexibility for irregular geographic domains compared to standard grid-based deep learning baselines.
Key Contributions
- Introduces CLDNet, a conditional latent neural ODE surrogate for shallow water equations that enables grid-independent urban flood modeling.
- Achieves a 115x speedup compared to GPU-accelerated solvers, reducing 96-hour simulations from ~55 minutes to ~29 seconds.
- Outperforms grid-based baselines (VAE-ConvLSTM, FNO) on irregular watershed benchmarks, attaining an 86% critical success index for inundation forecasting.
Open Questions & Future Work
Key Concepts
- cldnet-framework: A neural ODE-based surrogate for hydrodynamic simulation that uses coordinate-based decoding to enable grid-independent, high-resolution flood forecasting.
Archivist Review
I have approved the CLDNet framework as a distinct, reusable architectural paradigm for physics-informed surrogate modeling. I have also approved the open question regarding operational digital twin integration, as it identifies a systemic challenge in transitioning from static inference to active, data-assimilated digital twins. Other candidates were rejected as either too specific to the implementation or insufficiently novel for a permanent knowledge vault.
Approved Concepts
- Conditional Latent Dynamics Network (CLDNet): The architecture innovatively decouples spatial grid requirements from hydrodynamic surrogate modeling, a significant shift for physics-informed machine learning on irregular domains.
Approved Open Questions
- Operational Digital Twin Integration: Bridging the gap between surrogate inference and real-time state estimation is necessary to transform surrogate models into functional, operational digital twins.
Rejected Candidates
- [open_question] Improving Discharge Field Accuracy (
improving-discharge-field-accuracy) - low_impact: While technically valid, this is a specific performance improvement task rather than a fundamental scientific bottleneck suitable for a permanent vault note.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.13761
- paper_id
- 2605.13761
- paper_source
- openalex
- domain
- time-series
- tags
- physics-informed-lstmneural-parameterized-markov-operator
- architectures
-
- datasets
-
- concept_slugs
- cldnet-framework
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-16T06:02:36Z