📄 other

Physics-Informed Neural Networks for solving 2D Shallow Water Equations: Accuracy Assessment Through Dambreak Benchmark Problems

This paper demonstrates that a unified Physics-Informed Neural Network framework, which embeds the 2D shallow water equations directly into its loss function, achieves high accuracy (R² > 0.98) across four complex dam-break benchmarks, establishing it as a viable mesh-free alternative to traditional solvers for rapid flood risk assessment.

Original authors: Abdelkader Bemmoussat, Noureddine Maref, Zakaria Mahfoud, Khaled Korichi

Published 2026-07-24

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Abdelkader Bemmoussat, Noureddine Maref, Zakaria Mahfoud, Khaled Korichi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict how a flood will move through a city after a dam breaks. In the world of physics and engineering, this is a bit like trying to predict the path of a chaotic, invisible river of water that rushes over hills, crashes into walls, and dries up as it goes. For decades, scientists have used "grid-based" computers to solve this. Think of these grids like a giant chessboard laid over a map; the computer calculates the water's movement square by square. But if the map has weird shapes, like a winding river or a city with narrow alleys, drawing that chessboard is a nightmare. It takes forever, and if the water moves too fast, the squares can't keep up, leading to messy, inaccurate predictions.

Enter a newer, smarter idea called "Physics-Informed Neural Networks" (PINNs). Instead of a rigid chessboard, imagine a flexible, invisible net that learns the rules of physics as it goes. It doesn't need to be drawn on a grid; it just knows that water must follow the laws of gravity and momentum. This paper asks a big question: Can this flexible, grid-free net actually predict real-world floods as well as the old, heavy chessboard computers? The authors aren't just guessing; they are putting this new method through a series of tough "obstacle courses" to see if it can handle shockwaves, dry land turning wet, and tricky hills without getting confused.

The Great Dam-Break Race: Grids vs. The Invisible Net

In this study, a team of researchers from Djillali Liabes University in Algeria decided to put a new kind of computer brain to the test. They wanted to see if a "Physics-Informed Neural Network" (PINN) could solve the Shallow Water Equations—the fancy math rules that describe how water flows in rivers, floods, and tsunamis—without using the traditional, rigid grids that engineers have relied on for years.

Think of the old way of solving these problems like trying to paint a picture using only square tiles. You have to fit the tiles together perfectly, and if the picture has a curve or a sharp corner, you end up with a jagged, pixelated mess. The new PINN method is more like painting with a smooth, continuous brush. It doesn't need tiles; it just learns the "rules of the game" (the physics) and draws the water's path directly.

The researchers set up four different "obstacle courses" to see if their smooth brush could handle the chaos of a dam breaking.

The Obstacle Courses

The Simple Rush (1D Dam-Break): First, they tested a straight channel. In one version, the water rushed into a pool of existing water (wet bed). In the other, it crashed onto dry land (dry bed). This is like watching a wave hit a calm lake versus a wave hitting a dry beach. The goal was to see if the AI could predict exactly where the water would stop and how fast it would go.
The Hilly Hurdle (Triangular Bump): Next, they added a triangular hill in the middle of the channel. When the dam broke, the water had to rush over the hill, creating a complex mix of fast and slow flows. This tested if the AI could handle changes in the ground shape.
The Sideways Splash (2D Partial Dam-Break): Then, they moved to a 2D world. Imagine a dam that only breaks in the middle, sending water shooting out sideways like a firehose. This tested if the AI could handle water spreading out in all directions, not just in a straight line.
The Funnel Twist (Venturi Channel): Finally, they used a channel that got narrow and then wide again, like a funnel. This created a mix of slow and super-fast water, testing if the AI could handle extreme changes in speed and pressure.

The Results: The Net Holds Up

The results were surprisingly good. The PINN model acted like a masterful surfer, riding the waves of physics with impressive accuracy.

Accuracy: In almost every test, the AI's predictions matched the real-world data and the old "chessboard" computer models with a score (called R²) between 0.982 and 0.997. On a scale where 1.0 is perfect, that is a near-flawless performance.
Precision: The average error in predicting how deep the water would be was tiny—less than 0.01 meters (about the width of a finger).
Handling the Hard Stuff: The model did a great job with the tricky parts. It correctly predicted the "shock waves" (the sudden wall of water) and the "rarefaction fans" (the spreading out of water). It even handled the "wet-dry front," which is the messy edge where water meets dry land, a place where many computer models usually fail.

However, the paper is honest about where the new method stumbles. The AI had a bit more trouble with the sharpest, most sudden changes, like the very tip of the wave hitting dry land or the water rushing over the peak of the triangular hill. In these specific spots, the error was slightly higher, though still within acceptable limits for engineering.

Why This Matters

The biggest win for this new method isn't just that it's accurate; it's that it's mesh-free. The researchers found that because the AI doesn't need a grid, it can easily handle weird shapes and complex boundaries without the headache of drawing thousands of tiny squares. It's like having a map that redraws itself instantly to fit the terrain, rather than forcing the terrain to fit the map.

But there's a catch. While the AI is great at predicting the result once it's learned, the process of learning takes a lot of time and computing power. The researchers noted that training the network took up to 20,000 rounds of practice (epochs). It's like a student who gets an A on the final exam but had to study for weeks straight to get there, whereas the old grid method might have solved the problem faster if the map was simple.

The Bottom Line

This paper suggests that Physics-Informed Neural Networks are a powerful new tool for simulating floods. They can handle 1D and 2D dam-break scenarios, wet and dry beds, and complex hills with high accuracy. They are particularly promising for areas with strange shapes where old methods struggle. However, they aren't a magic wand yet; they are currently slower to train and still struggle a tiny bit with the sharpest, most chaotic edges of a flood. The authors conclude that while this isn't a replacement for every old method today, it is a very strong contender for the future of flood prediction, especially when the terrain gets complicated.

Technical Summary: Physics-Informed Neural Networks for Solving 2D Shallow Water Equations

Problem Statement
Accurate simulation of free-surface flows, particularly dam-break scenarios, is critical for flood risk assessment and engineering design. These problems involve complex physical phenomena, including supercritical–subcritical transitions, moving wet–dry fronts, hydraulic jumps, and interactions with variable bathymetry. While classical grid-based solvers (Finite Difference, Finite Element, and Finite Volume Methods) are mature, they face structural limitations: mesh generation for complex or evolving domains is time-consuming, resolving sharp flow features requires computationally expensive fine grids, and explicit schemes are constrained by the CFL stability condition. Conversely, purely data-driven machine learning models lack guarantees of satisfying physical conservation laws and often fail to extrapolate outside their training distributions. There is a need for a mesh-free, physics-constrained approach that can handle the nonlinear hyperbolic nature of the Shallow Water Equations (SWEs) across both 1D and 2D domains without prohibitive computational costs.

Methodology
The authors propose a unified Physics-Informed Neural Network (PINN) framework to solve the 2D SWEs. The core methodology involves embedding the governing partial differential equations (PDEs) directly into the neural network's training process via a composite loss function.

Governing Equations: The framework solves the depth-averaged 2D SWEs in conservative form, accounting for continuity and momentum equations with bed slope and friction source terms (Manning–Strickler).
Network Architecture: A deep neural network with fully connected hidden layers uses the hyperbolic tangent (tanh) activation function. The network takes spatial-temporal coordinates $(x, y, t)$ as input and outputs the conserved variables: water depth ( $h$ ) and momentum components ($hu, hv$). To improve accuracy in multivariate problems, sub-networks for each output variable receive the same raw inputs but maintain decoupled trainable parameters.
Loss Function: The training minimizes a weighted sum of Mean Squared Errors (MSE) comprising:
- PDE Residuals: Enforced at interior collocation points using automatic differentiation to compute exact partial derivatives, ensuring the solution satisfies the continuity and momentum equations.
- Initial Conditions (ICs) and Boundary Conditions (BCs): Enforced at specific points to match the initial state and domain boundaries.
Training Strategy: A two-phase optimization process is employed:
1. Global Exploration: The Adam optimizer reduces the loss rapidly over 20,000–50,000 epochs.
2. Local Refinement: The L-BFGS algorithm fine-tunes weights to achieve near machine-precision residuals.
- Collocation Strategy: Latin Hypercube Sampling (LHS) ensures uniform distribution of points, with adaptive resampling applied every 5,000 iterations to concentrate points in regions of high PDE residuals (e.g., shock waves and wet–dry fronts).

Benchmark Cases
The framework was rigorously validated against four benchmark problems of increasing complexity, comparing PINN predictions against analytical solutions and experimental data:

1D Ideal Dam-Break (Wet and Dry Beds): Tests shock wave and rarefaction fan capture.
1D Dam-Break over a Triangular Hump: Evaluates handling of variable bathymetry, wet–dry fronts, and hydraulic jumps (CADAM project data).
2D Partial Dam-Break: Assesses lateral spreading and shock interactions in a 2D domain (Fraccarollo and Toro experiment).
2D Dam-Break in a Venturi Channel: Tests performance in convergent-divergent geometries with mixed flow regimes (Franco experiment).

Key Results
The PINN framework demonstrated high fidelity across all configurations:

Accuracy: The model achieved $R^2$ values between 0.982 and 0.997 across all test cases. Root Mean Square Errors (RMSE) for water depth were consistently below 0.01 m, and $L_2$ errors remained within engineering tolerances.
Physical Consistency: The model successfully reproduced shock wave propagation, rarefaction fans, wet–dry front advancement, and flow acceleration through channel contractions without explicit upwinding or Riemann solvers.
Specific Challenges:
- In the 1D dry-bed case, velocity errors were slightly higher due to the singularity at the advancing wet–dry front, though free-surface predictions remained accurate.
- In the 2D Venturi channel, the model handled complex pressure gradients and shock reflections effectively, showing superior $L_2$ accuracy compared to other metrics in constrained geometries.
- Minor discrepancies were noted at specific gauge points located at obstacle crests or in highly dynamic 3D-like flow zones (e.g., near the breach in the partial dam-break), where depth-averaged assumptions limit resolution.

Significance and Claims
The paper positions this unified PINN framework as a practical surrogate for finite-volume solvers in rapid flood risk assessment, particularly for geometrically complex domains. The authors highlight four specific advantages:

Mesh-Free Flexibility: Eliminates the need for grid generation, which is beneficial for irregular geometries like the Venturi channel.
Physical Consistency: Enforces conservation laws directly through the loss function, removing the need for numerical flux limiters or Riemann solvers.
Differentiability: Provides smooth, differentiable solution fields suitable for post-processing tasks like sensitivity analysis.
Query Efficiency: Once trained, the model can evaluate solutions at arbitrary query points without re-running the full system.

Limitations and Future Directions
The authors modestly acknowledge current limitations:

Training Cost: The non-convex, multi-objective optimization is computationally expensive, requiring up to 20,000 epochs and exceeding the runtime of a single forward pass of an optimized FVM solver.
Discontinuity Resolution: Sharp gradients and wet–dry interfaces still produce elevated errors, indicating challenges in representing discontinuities within a single smooth network.
Depth-Averaged Assumption: The model is limited by the hydrostatic assumption, which may not hold near breaches where vertical accelerations are significant.

The paper concludes that while PINNs offer a physically grounded, mesh-free path for simulating dam-break flows with engineering-grade accuracy, future work must address training costs and sharp-gradient resolution through advanced optimization strategies, adaptive architectures, and potential hybrid coupling with established solvers to enable operational flood forecasting.