How Sparse and How Noisy? Systematic Benchmarking of Inverse Physics-Informed Neural Networks for Manning Friction Estimation in Shallow Water Equations

This study systematically benchmarks the reliability of inverse Physics-Informed Neural Networks for estimating Manning friction in shallow water equations, revealing that while two-dimensional flows allow robust recovery with sparse and noisy data, one-dimensional flows suffer from structural identifiability limitations and that joint depth-velocity observations are essential for accurate parameter identification.

The Gist

The Big Picture: Teaching a Robot to "Feel" the River

Imagine you are trying to figure out how rough the bottom of a river is (how much the rocks and mud slow the water down). In engineering, this is called the Manning friction coefficient. Usually, you'd need to go out, measure the riverbed everywhere, and do complex math to guess this number.

This paper tests a new kind of "smart robot" (called a Physics-Informed Neural Network, or PINN). This robot has a special superpower: it knows the laws of physics (how water should flow) but it has to learn the specific details of this river by looking at very few, very messy data points.

The researchers asked: "How many data points do we need, and how messy can the data be, before this robot stops guessing correctly?"

They tested the robot on two different "river" scenarios: a simple straight channel (1D) and a more complex, curved channel (2D).

---

The Experiment: The "Blindfolded Detective" Game

Think of the robot as a detective trying to solve a crime (finding the friction value) based on clues left at the scene.

The Clues (Observations):
The robot is given measurements of water depth and speed at specific spots.

• Sparsity: Sometimes the detective only gets 5 clues (very sparse). Other times, they get 50 clues (dense).
• Noise: Sometimes the clues are perfect. Other times, the clues are "noisy"—like someone whispering the wrong numbers or the tape measure being slightly broken (up to 20% error).
• The Type of Clue: Sometimes the detective only sees the water depth. Sometimes they only see the speed. Sometimes they see both.

The Goal:
The robot must guess the "roughness" of the riverbed. If it guesses correctly, it can predict how the whole river flows, even in places where it didn't take a measurement.

---

The Results: What Worked and What Didn't

1. The Simple River (1D Benchmark)

Imagine a straight, narrow canal.

• The Problem: Even when the robot had perfect data, it kept guessing the river was about 15% rougher than it actually was.
• The Analogy: It's like a detective who is so convinced the suspect is guilty that they ignore the evidence and stick to their wrong theory. No matter how many more clues they got (5, 10, or 50), the mistake didn't get better.
• The Lesson: In this simple setup, the robot hit a "structural wall." It couldn't figure out the roughness just by looking at the water level or speed alone. It needed both depth and speed together to even start guessing, but even then, it had a built-in bias.

2. The Complex River (2D Benchmark)

Imagine a wide river with a curved bottom, where water swirls sideways a little bit.

• The Success: This was much easier for the robot!
  • The "Sweet Spot": If the robot had just 10 clues (measurements of depth and speed), it guessed the roughness with less than 5% error.
  • The Noise Tolerance: If the clues were a bit messy (up to 10% noise), the robot still did a great job. Even with 50 clues and very messy data (20% noise), it stayed accurate.
  • The Danger Zone: If the robot only had 5 clues, it got confused and gave wildly different answers every time it tried. It was too little information to work with.
• Why it worked better: The extra "swirl" in the water (sideways velocity) gave the robot a new clue. It was like the detective getting a fingerprint and a shoe print, instead of just a shoe print. The sideways movement told the robot exactly how rough the bottom was.

3. The "One-Clue" Trap

The researchers tested what happens if the robot only sees water depth (no speed) or only speed (no depth).

• The Result: Total failure. The robot guessed the roughness was either almost zero or huge.
• The Lesson: You cannot solve the puzzle with just one type of clue. You need the full picture (depth + speed) to make sense of the friction.

---

The Secret Sauce: How the Robot Learned

The researchers found that how they taught the robot mattered just as much as the data.

1. The Two-Step Dance:

   • Step 1: First, they told the robot to ignore the physics laws and just memorize the messy data points. This helped it get a general idea of what the river looked like.
   • Step 2: Then, they turned on the "physics laws" to correct the guess.
   • Why? If they turned on the physics laws immediately, the robot would cheat. It would say, "I'll just make the river friction zero so the math is easy," and the whole thing would break. The two-step process prevented this cheating.

2. The "Second-Order" Trap:

   • Standard AI training often uses a "fine-tuning" step at the end to polish the answer. The researchers found that for this specific problem, polishing made it worse.
   • Analogy: Imagine a sculptor who has a rough clay statue. If they try to use a super-precise laser to smooth it out, they might accidentally carve away the wrong parts because the clay is too soft. The robot got "too precise" and fell into a trap where it found a mathematically perfect but physically wrong answer. They had to stop the robot before it got too fancy.

Summary: The Takeaway

• Can we use AI to guess river friction? Yes, but only if we have the right kind of data.
• How much data? You need at least 10 good measurement spots. Five is not enough.
• How messy can the data be? If you have 50 spots, the data can be quite messy. If you only have 10 spots, the data needs to be fairly clean.
• What kind of data? You must measure both how deep the water is and how fast it's moving. Measuring just one is useless.
• The Catch: In very simple, straight rivers, the AI might still be slightly biased (wrong by a fixed amount) no matter how much data you give it. But in complex, real-world rivers with curves, it works very well.

This paper essentially built a "rulebook" for engineers: Here is exactly how many sensors you need and how clean your data must be before you can trust this AI to tell you how rough a river is.

Technical Summary

Technical Summary: Systematic Benchmarking of Inverse PINNs for Manning Friction Estimation

Problem Statement
Physics-informed neural networks (PINNs) offer a framework for inverse hydrodynamic modeling by combining sparse observations with governing physical constraints. However, their reliability for estimating hydraulic parameters, specifically the Manning friction coefficient ($n$) in the shallow water equations (SWEs), under realistic data limitations (sparsity and noise) remains insufficiently characterized. Existing studies often demonstrate successful recovery in specific, idealized cases but lack a systematic quantification of how estimation error varies with observation density, noise magnitude, and variable type. This gap is critical because practical coastal and flood modeling applications rarely possess dense, noise-free observations, and inverse problems involving friction are prone to equifinality, where different parameter combinations produce similar water-level responses.

Methodology
The study establishes a reproducible benchmarking framework to evaluate inverse PINN recovery of the Manning coefficient under controlled variations in observation sparsity, noise, and type.

• Governing Equations: The framework utilizes the conservative 1D and 2D shallow water equations with Manning bottom friction. The friction slope is defined as $S_f = n^2 u |u| / h^{4/3}$ (1D) and extended to vector components in 2D.
• Benchmark Cases:
  1. 1D Case: A subcritical channel based on the MacDonald analytical solution. The bed elevation is back-solved to ensure a self-consistent steady state for a target $n_{true} = 0.02$.
  2. 2D Case: A sloped channel with a parabolic transverse bed. The reference steady state is generated using a well-balanced finite-volume solver (HLL Riemann solver, Audusse hydrostatic reconstruction) with a target $n_{true} = 0.02$.
• Inverse PINN Formulation: A fully connected feed-forward network (8 hidden layers, 20 neurons, tanh activation) maps spatial coordinates to flow variables ($h, u, v$). The Manning coefficient is treated as a trainable positive scalar, parameterized via softplus to enforce positivity.
• Training Strategy: A two-phase optimization is employed:
  • Phase A: The network trains exclusively on observation loss ($\lambda_{eq}=0$) to establish a flow field consistent with data.
  • Phase B: The physics residual is activated ($\lambda_{eq}=1.0$), and the network and friction parameter are jointly optimized.
  • Optimizer: Adam optimizer is used exclusively. Diagnostic experiments showed that adding a second-order L-BFGS refinement step degraded the friction recovery by exploiting near-null directions in the parameter landscape.
• Experimental Design: The benchmark sweeps across:
  • Sparsity: $N_{obs} \in \{5, 10, 20, 50\}$.
  • Noise: Gaussian noise with standard deviation $\sigma \in \{0\%, 5\%, 10\%, 20\%$ of field standard deviation.
  • Observation Type: In 1D, ablation studies isolate depth ($h$), velocity ($u$), and combined ($h, u$) observations.
  • Statistics are aggregated over 5 independent random seeds to distinguish robust regimes from non-identifiable configurations.

Key Results

• Two-Dimensional Benchmark (Robust Recovery):

  • The 2D case achieves robust friction recovery with relative errors below 5% when at least 10 observations of depth and velocity are available and noise levels are $\le 10\%$.
  • A clear identifiability threshold exists: recovery is unstable at $N_{obs}=5$ (error $\approx 22\%$ with high variance) but stabilizes abruptly at $N_{obs}=10$.
  • With 50 observations, the method remains robust up to 20% noise.
  • The 2D case consistently outperforms the 1D case (errors $\approx 4\%$ vs. $\approx 15\%$), attributed to the information content of the transverse velocity component, which scales directly with friction via the cross-momentum equation.

• One-Dimensional Benchmark (Structural Limitation):

  • The 1D case exhibits a persistent positive bias of approximately 15% relative to the true value ($n_{true}=0.02$).
  • This bias is largely insensitive to observation count (ranging from 5 to 50) and noise levels (up to 20%), indicating a structural identifiability limitation rather than a data-density issue.
  • Observation-Type Ablation: Joint observation of depth and velocity is essential. Recovery fails catastrophically with depth-only observations (error $\approx 96\%$) and velocity-only observations (error $\approx 62\%$), regardless of sampling density.

• Training Dynamics:

  • The two-phase training schedule is critical; activating physics residuals immediately leads to the friction parameter collapsing toward zero.
  • Implementation details regarding loss weights (mutable state vs. compile-time constants) were found decisive for gradient flow to the friction parameter.

Significance and Claims
The paper claims to provide a reproducible benchmark that explicitly quantifies the trade-off between sensor density, noise magnitude, and friction-recovery accuracy. It does not propose a universal calibration law but offers an interpretable reference for determining when inverse PINNs can and cannot reliably estimate Manning friction.

Key contributions include:

1. Identifiability Thresholds: Defining specific conditions (e.g., $N_{obs} \ge 10$ in 2D, joint $h$-$u$ observation) required for reliable inverse identification.
2. Structural vs. Data Limitations: Distinguishing between regimes where error is driven by data scarcity (2D low-density) versus intrinsic structural limitations (1D bias).
3. Methodological Corrections: Highlighting that second-order refinement (L-BFGS) can be detrimental for inverse PINNs and that specific implementation choices in automatic differentiation frameworks can silently prevent parameter updates.
4. Practical Guidance: Demonstrating that for inverse SWE problems, co-located depth and velocity measurements are significantly more informative than dense measurements of a single variable, a finding with direct implications for observation network design in coastal engineering and hydrology.