A Neural Operator Emulator for Coastal and Riverine Shallow Water Dynamics

This paper introduces MITONet, a novel neural operator emulator that achieves real-time, high-accuracy forecasting of complex coastal and riverine shallow water dynamics with significant computational speedups (100x–1,250x) and robust generalization to unseen conditions and parameters.

The Gist

Imagine you are trying to predict how water moves through a complex network of rivers, bays, and inlets during a storm or a daily tide. Traditionally, scientists use massive, super-computer-powered simulations to do this. Think of these simulations like a high-end video game engine: they are incredibly accurate, calculating every single ripple and current, but they are slow. Running a simulation for a whole month might take hours or even days of computing time. This is too slow if you need a quick answer for emergency planning or daily decisions.

On the other hand, there are simpler, faster methods, but they are like using a blurry, low-resolution map. They are quick, but they often get lost when the weather changes or the water behaves in a new way. They struggle to predict what happens in situations they haven't seen before.

The Solution: MITONet
The authors of this paper introduce a new tool called MITONet. You can think of MITONet as a "super-smart student" that has studied thousands of hours of high-quality water simulations. Instead of trying to calculate every single drop of water from scratch every time (like the slow super-computer), MITONet has learned the rules of how the water behaves.

Here is how it works, using some everyday analogies:

1. The "Compression" Trick (The Autoencoder):
   Imagine you have a giant, detailed 3D model of a city. It's too big to carry around. MITONet first learns to shrink this giant model down into a tiny, compact "blueprint" or a "latent space" (like a highly compressed zip file). It learns to see the big picture without getting bogged down in every tiny detail. This makes the math much faster.

2. The "Multiple Inputs" (The Branches):
   Water doesn't just move because of one thing. It moves because of the starting water level, the wind, the tides, and how rough the riverbed is (like mud vs. smooth rock). MITONet has special "branches" in its brain that look at each of these factors separately. It's like having a team of experts: one looks at the wind, one looks at the riverbed, and one looks at the starting water level. They all talk to each other to figure out the next step.

3. The "Time-Travel" Trick (Temporal Bundling):
   Usually, when you predict the future step-by-step (like predicting tomorrow, then the day after, then the day after), small mistakes pile up, and by day 100, your prediction is totally wrong. MITONet uses a trick called "temporal bundling." Instead of taking one tiny step at a time, it learns to jump ahead in chunks (like taking 5 steps at once). This keeps the prediction stable and accurate for a very long time, even up to 175 days into the future.

What Did They Test?
The team tested this "student" on two very different real-world scenarios:

• Shinnecock Inlet, New York: A coastal area where the ocean tides push water in and out. This is a rhythmic, predictable dance.
• Red River, Louisiana: A river with a chaotic, changing flow where water rushes in from upstream and pushes out downstream. This is a messy, unpredictable rush.

The Results
MITONet was amazing at both.

• Speed: It was 100 to 1,250 times faster than the traditional super-computer simulations. A task that took the super-computer hours took MITONet just seconds.
• Accuracy: Even when they asked it to predict water levels for conditions it had never seen before (like a new type of riverbed roughness or a completely random starting point), it was still incredibly accurate. It got the "shape" of the water movement right more than 90% of the time.
• Stability: It didn't get confused or drift off course even after predicting 175 days into the future.

The Catch
The paper notes one limitation: MITONet is like a student who knows the map of a specific city perfectly but can't instantly draw a map for a different city it has never seen. It works great for the specific shapes of the Shinnecock Inlet and the Red River, but it can't magically predict water flow in a completely new, unseen geography without retraining.

In Summary
MITONet is a new, lightning-fast tool that learns the physics of water movement from data. It acts like a "neural emulator," giving us the accuracy of a slow, expensive super-computer simulation but with the speed of a simple calculation. This means we can get real-time, accurate predictions for floods and tides much faster, helping us plan and react to extreme weather events more effectively.

Technical Summary

Technical Summary: A Neural Operator Emulator for Coastal and Riverine Shallow Water Dynamics

Problem Statement
Coastal regions and river floodplains are highly vulnerable to extreme weather, necessitating accurate real-time forecasting of hydrodynamic processes for infrastructure planning and climate adaptation. High-fidelity numerical models based on the two-dimensional shallow water equations (2D SWE), such as ADCIRC and AdH, provide accurate simulations but are often computationally prohibitive for real-time applications. Conversely, traditional model order reduction and standard neural networks often struggle to generalize to out-of-distribution conditions, such as varying boundary conditions, domain parameters, or initial states. While recent advances in neural operators (e.g., DeepONet, FNO) have shown promise, existing applications in hydrodynamics often face limitations: they may rely solely on initial conditions without incorporating time-varying boundary forcings, fail to handle varying domain parameters, or predict only specific quantities of interest (e.g., inundation extent) rather than the full space-time evolution of state variables.

Methodology: MITONet Framework
The authors introduce MITONet (Multiple-Input Temporal Operator Network), an autoregressive neural operator framework designed to emulate high-dimensional numerical solvers for complex, nonlinear, time-dependent, and parameterized PDEs. The architecture integrates four key strategies to address scalability, generalizability, and stability:

1. Latent Space Learning: To handle high-dimensional data efficiently, MITONet employs a pre-trained autoencoder network to map the high-dimensional solution snapshots (water surface elevation $H$ and velocity components $U, V$) into a lower-dimensional latent space. The neural operator learns the mapping within this latent space, significantly reducing computational complexity.
2. Multiple-Input Handling: Unlike standard DeepONet variants that often concatenate inputs, MITONet utilizes multiple independent branch networks to separately encode distinct input functions. These inputs include:
   • Initial Conditions (IC) in the latent space.
   • Time-varying Boundary Conditions (BC) (e.g., tidal elevations, inflow discharge, tailwater elevation).
   • Domain parameters (specifically the scalar bottom friction coefficient, $r$).
     A trunk network encodes the temporal coordinates of the output.
3. Enhanced Information Propagation: To mitigate the vanishing gradient problem and improve expressivity, the framework incorporates interconnected "encoder" networks (following the M-DeepONet concept) for both branch and trunk networks. These encoders facilitate the exchange of information between the multiple input branches and the trunk network during the forward pass.
4. Temporal Bundling: To enable long-horizon autoregressive forecasting and minimize error accumulation, the model is trained using temporal bundling. Instead of predicting a single time step, the network predicts a window of $\tau$ future steps synchronously. During inference, the prediction at the end of the window serves as the initial condition for the next autoregressive step.

Experimental Setup
The framework was evaluated on two distinct, realistic hydrodynamic problems governed by the 2D SWE:

• Shinnecock Inlet (NY): A coastal circulation problem driven by time-dependent tidal boundary conditions. The domain features complex geometries and variable mesh resolutions. The study varied the bottom friction coefficient ($r$) across a wide range (0.0025 to 0.1) to test parametric extrapolation.
• Red River (LA): A riverine flow problem characterized by non-periodic, time-varying inflow and tailwater elevation boundary conditions. The domain involves an unstructured triangular mesh and realistic bathymetry. The bottom friction coefficient was varied within a narrower, more natural range (0.02375 to 0.02625).

In both cases, the models were trained on high-fidelity numerical simulation data (generated by ADCIRC and AdH, respectively) and tested on unseen parameter values and initial conditions, including random and zero initial states.

Key Results

• Predictive Skill: MITONet demonstrated superior performance compared to vanilla DeepONet, Modified DeepONet, MIONet, and Latent DeepONet. In the Shinnecock Inlet case, MITONet achieved anomaly correlation coefficients (ACC) above 0.9 and a maximum normalized root mean square error (NRMSE) of 0.011 across 175-day autoregressive rollouts.
• Stability and Long-Term Forecasting: The framework maintained high accuracy over long forecast horizons (up to 175 days or 8,400 time steps) with minimal error accumulation. For the Shinnecock example, the RMSE increased by less than 9% when extending the forecast from 55 to 175 days.
• Generalization: The model successfully generalized to unseen bottom friction coefficients (parametric extrapolation) and unseen initial conditions, including "hotstarts" (random ICs) and "coldstarts" (zero ICs), where the model ramped up to accurate predictions within approximately two days.
• Computational Efficiency: MITONet provided substantial speed-ups over the physics-based solvers. In the Shinnecock case, it achieved a 109x speed-up over ADCIRC. In the Red River case, it achieved a 1,125x speed-up over AdH, generating 60-day forecasts in seconds on a single GPU.
• Versatility: The framework successfully emulated two distinct flow regimes (quasi-periodic tidal dynamics and aperiodic riverine dynamics) with different geometries, mesh resolutions, and driving forces, demonstrating its flexibility.

Significance and Claims
The paper claims that MITONet represents a significant advancement in data-driven hydrodynamic modeling by fusing latent-space operator learning, multi-input handling, and temporal bundling into a structured framework. Its primary significance lies in its ability to serve as a general hydrodynamic emulator rather than a task-specific surrogate. Unlike previous studies that focused on predicting specific metrics (e.g., flood extent or gauge readings) or relied on initial conditions alone, MITONet learns the full space-time evolution of all shallow-water state variables ($H, U, V$) while explicitly incorporating time-varying boundary conditions and domain parameters.

The authors emphasize that while the framework offers remarkable speed and accuracy, it is not a replacement for physics-based models in scenarios involving extreme events far outside the training distribution. They acknowledge current limitations, specifically the lack of discretization invariance (the inability to spatially interpolate solutions to arbitrary points in the domain due to the fixed latent space representation) and the reliance on specific mesh geometries. The paper concludes that MITONet provides a robust foundation for real-time planning and decision-making in coastal and riverine environments, with future work needed to address geometry transfer, uncertainty quantification, and the integration of physics-based residual losses.