Physics-informed neural operators for the in situ characterization of locally reacting sound absorbers

This paper introduces a physics-informed neural operator framework that robustly estimates frequency-dependent surface admittance of locally reacting sound absorbers directly from noisy near-field measurements by embedding governing acoustic equations into the training process, thereby enabling accurate in situ characterization without explicit forward modeling.

The Gist

The Big Picture: The "Acoustic Detective" Problem

Imagine you are a sound engineer trying to design a concert hall or a quiet car. To do this perfectly, you need to know exactly how sound bounces off or gets absorbed by the walls, seats, or dashboard. In physics, this property is called acoustic admittance (or impedance).

The Problem:
Usually, to find out how a material absorbs sound, you have to take a small piece of it to a lab and put it in a special tube (like a giant straw) to test it. But in the real world, you can't always cut a piece of the wall out of a building or a car. You need to measure it right there (in situ) without touching it.

The Challenge:
When you stand in front of a wall and measure the sound, you only hear the "echo" and the "vibration" near the surface. Figuring out the exact properties of the wall from these measurements is like trying to guess the ingredients of a soup just by tasting a spoonful while it's boiling. It's a messy math problem called an "inverse problem." If there is even a little bit of background noise (like a car honking outside), traditional math methods often get confused and give you the wrong answer.

The Solution: The "Smart Physics Detective"

The authors of this paper created a new kind of AI detective called a Physics-Informed Neural Operator. Here is how it works, broken down into simple concepts:

1. The Old Way vs. The New Way

• The Old Way (Pure Data): Imagine teaching a student to guess the soup ingredients by feeding them 1,000 photos of soups and their recipes. If you show them a photo of a soup they've never seen before, they might guess wrong because they just memorized patterns, not the rules of cooking.
• The New Way (Physics-Informed): Now, imagine teaching that same student, but you also give them a cookbook that explains the laws of cooking (e.g., "if you add salt, it gets salty"). If the student guesses "sugar" for a salty soup, the cookbook corrects them immediately. They learn both the patterns and the rules.

This paper's AI does exactly that. It learns from data (measurements) but is constantly corrected by the Laws of Physics (specifically, how sound waves move and bounce).

2. The "DeepONet" (The Super-Brain)

The AI they built is called a Deep Operator Network (DeepONet).

• Analogy: Think of a standard AI as a calculator that solves one specific math problem at a time. If you change the frequency (pitch) of the sound, you have to teach the calculator a whole new lesson.
• The DeepONet: This is like a master chef who has learned the concept of cooking. Once trained, they can instantly cook a dish for any ingredient list or any temperature without needing a new recipe book. In this case, the AI learns the relationship between sound measurements and the wall's properties across all frequencies at once. It doesn't need to relearn anything if you change the pitch of the sound.

3. How It Works (The Training Process)

The researchers didn't just throw random data at the AI. They built a "training gym" for it:

1. The Simulator: They used a super-computer to create a fake, perfect world where they knew exactly what the sound-absorbing material was.
2. The Messy Data: They added "noise" (static) to the measurements to make it look like a real, imperfect recording.
3. The Physics Rules: They told the AI: "You must obey the laws of sound."
   • The Momentum Rule: If the air moves one way, the pressure must change in a specific way.
   • The Wave Rule: Sound waves must ripple through the air correctly.
   • The Boundary Rule: When the wave hits the wall, it must react according to the wall's properties.

If the AI guesses a wall property that breaks these rules, it gets a "penalty" (a bad grade). This forces the AI to find the answer that fits the data and the laws of physics simultaneously.

The Results: Why It's a Game Changer

The team tested this on two types of foam (Melamine and Polyurethane) used for soundproofing.

• Accuracy: The AI correctly identified the sound-absorbing properties of the foam across a wide range of pitches, from low rumbles to high squeaks.
• Noise Resistance: This is the big win. When they added a lot of static noise to the data, the old "pure math" methods failed completely. The "Physics-Informed" AI, however, ignored the noise and stuck to the physical laws, giving the correct answer.
• Sparse Data: Even when they gave the AI very few measurement points (like having only a few microphones instead of a whole array), it still worked well because the physics rules filled in the gaps.

The Bottom Line

This paper introduces a tool that acts like a super-smart, physics-savvy detective. Instead of needing expensive, perfect lab equipment to measure sound-absorbing materials, you can now use a few microphones and this AI to figure out exactly how a material behaves in the real world, even if the environment is noisy or the data is incomplete.

It's like upgrading from a magnifying glass to a X-ray vision that understands the rules of the universe, allowing us to "see" the hidden acoustic properties of materials right where they are installed.

Technical Summary

1. Problem Statement

Accurate knowledge of acoustic surface admittance (or impedance) is critical for reliable wave-based simulations (e.g., FEM, BEM) of acoustic fields. However, estimating these properties in situ (in real-world conditions) remains challenging due to:

• Limitations of Standard Methods: Impedance tubes (ISO 10534-25) are restricted to normal incidence and idealized mounting, while reverberation room methods (ISO 354) assume diffuse fields and only provide magnitude data, neglecting crucial phase information.
• Inverse Problem Instability: Traditional in situ methods often treat the problem as an ill-posed inverse problem. They rely on explicit forward models (analytical or numerical) and are highly sensitive to measurement noise, modeling errors (e.g., edge diffraction), and sparse data.
• Limitations of Existing ML Approaches:
  • Physics-Informed Neural Networks (PINNs): While effective, standard PINNs require retraining for every new frequency, making them inefficient for frequency-dependent phenomena.
  • Purely Data-Driven Models: These lack physical consistency and often fail to generalize well when data is noisy or sparse.

2. Methodology

The authors propose a Physics-Informed Deep Operator Network (DeepONet) to estimate frequency-dependent surface admittance directly from near-field measurements of sound pressure ($p$) and normal particle velocity ($v_z$).

A. Data Generation & Setup

• Simulation Model: A Boundary Element Method (BEM) model simulates a finite, planar porous sample (50 cm $\times$ 50 cm) on a rigid baffle under semi-free-field conditions.
• Excitation: A monopole source excites the system.
• Sensors: A double-layer planar array (13 $\times$ 13 sensors per layer) measures $p$ and $v_z$ at discrete points.
• Reference Data: Two materials (Melamine foam and PU foam) are characterized using impedance tube measurements to generate ground-truth frequency-dependent admittance spectra ($Y_{ref}$).
• Noise: Gaussian noise is added to simulate realistic measurement conditions (SNR = 40 dB).

B. Deep Operator Network Architecture

Unlike standard neural networks that map vectors to vectors, DeepONets learn mappings between infinite-dimensional function spaces.

• Branch Network: Encodes the input frequency function (discretized frequency samples).
• Trunk Network: Encodes spatial coordinates ($x, y, z$).
• Output: The networks combine via an inner product to predict the acoustic field quantities ($p_\theta, v_{z,\theta}$) and the global surface admittance ($Y_\theta$) simultaneously.
• Advantage: Once trained, the model can predict fields for unseen frequencies and spatial coordinates without retraining, solving the "frequency-wise inversion" bottleneck.

C. Physics-Informed Loss Function

The training objective minimizes a weighted sum of data-driven and physics-based losses:
$$L(\theta) = \lambda_p L_p + \lambda_{vz} L_{vz} + \lambda_{LM} L_{LM} + \lambda_{Helm} L_{Helm} + \lambda_R L_R$$

1. Data Loss ($L_p, L_{vz}$): Mean Squared Error (MSE) between predictions and noisy measurement data.
2. Linearized Momentum Loss ($L_{LM}$): Enforces the relationship $\frac{\partial p}{\partial z} = -i\omega\rho_0 v_z$.
3. Helmholtz Loss ($L_{Helm}$): Enforces the wave equation in the domain. A mixed formulation is used to reduce derivative order, improving numerical stability.
4. Robin Boundary Loss ($L_R$): Enforces the boundary condition at the sample surface: $i\omega\rho_0 v_z + ikY(f)p = 0$.
   • Key Innovation: The surface admittance $Y_\theta(f)$ is treated as a trainable parameter optimized jointly with the network weights. This allows the model to infer a globally consistent admittance spectrum that minimizes the boundary residual across the entire surface, rather than calculating point-wise ratios.

3. Key Contributions

1. Operator Learning for Acoustics: First application of DeepONets to infer frequency-dependent surface admittance, enabling a single model to handle the entire frequency spectrum without retraining.
2. Global Admittance Inference: Instead of estimating admittance point-by-point (which is noise-sensitive), the method infers a spatially uniform admittance spectrum as a global parameter, significantly improving robustness.
3. Physics-Based Regularization: By embedding the Helmholtz equation, momentum equation, and Robin boundary conditions directly into the loss function, the method achieves physical consistency and noise robustness without requiring an explicit forward solver during inference.
4. Elimination of Forward Models: The approach learns the mapping from measurements to physical quantities directly, bypassing the need for iterative optimization or explicit analytical forward models that introduce modeling errors.

4. Results

The method was validated on synthetic data for Melamine and PU foams across 100–5000 Hz.

• Accuracy:
  • Admittance: Achieved low relative $L_2$-norm errors (Mean $\mu \approx 0.027$ for Melamine, $0.028$ for PU).
  • Field Quantities: Sound pressure predictions were highly accurate ($\mu \approx 0.019-0.022$). Particle velocity showed slightly higher errors at low frequencies due to weak gradients but remained robust.
  • Spatial Consistency: Modal Assurance Criterion (MAC) values were near unity ($>0.98$), indicating excellent agreement in both amplitude and phase structure.
• Robustness to Noise:
  • Even at low SNR (12 dB), the physics-informed model maintained a mean error of 0.165, whereas the purely data-driven variant failed significantly.
  • The physics-informed approach showed superior performance even with sparse measurement data (down to 18 points), whereas data-only models required dense sampling to converge.
• Frequency Range: The model successfully captured complex frequency-dependent behaviors, including peaks and reactive/dissipative transitions, across more than 5.5 octaves.

5. Significance and Conclusion

This work presents a paradigm shift in acoustic material characterization:

• Efficiency: It eliminates the computational cost of frequency-wise inversion and iterative forward simulations.
• Robustness: It solves the ill-posed nature of inverse problems by using physical laws as strong regularizers, making it viable for noisy, sparse, real-world measurements.
• Scalability: The operator learning framework allows for rapid prediction of acoustic fields and material properties for new frequencies and spatial configurations.

Future Work: The authors plan to validate the method with real experimental data (moving beyond synthetic noise) and extend the framework to complex, non-free-field environments (e.g., rooms with strong reflections and modal behavior).

In summary, the proposed Physics-Informed Deep Operator Network offers a powerful, efficient, and robust alternative to conventional in situ characterization methods, capable of delivering physically consistent acoustic material properties directly from near-field measurements.