Surface impedance inference via neural fields and sparse acoustic data obtained by a compact array

Standardized laboratory characterizations for absorbing materials rely on idealized sound field assumptions, which deviate largely from real-life conditions. Consequently, \emph{in-situ} acoustic characterization has become essential for accurate diagnosis and virtual prototyping. We propose a physics-informed neural field that reconstructs local, near-surface broadband sound fields from sparse pressure samples to directly infer complex surface impedance. A parallel, multi-frequency architecture enables a broadband impedance retrieval within runtimes on the order of seconds to minutes. To validate the method, we developed a compact microphone array with low hardware complexity. Numerical verifications and laboratory experiments demonstrate accurate impedance retrieval with a small number of sensors under realistic conditions. We further showcase the approach in a vehicle cabin to provide practical guidance on measurement locations that avoid strong interference. Here, we show that this approach offers a robust means of characterizing \emph{in-situ} boundary conditions for architectural and automotive acoustics.

💡 Research Summary

The paper addresses a critical gap in acoustic material characterization: the inability of standardized laboratory methods (e.g., ISO 354, impedance tube) to capture the true in‑situ behavior of absorbers, which is essential for high‑fidelity virtual prototyping in architecture and automotive acoustics. While absorption coefficients are easy to measure, they are non‑unique with respect to the underlying complex surface impedance, a physically meaningful quantity that directly governs boundary conditions. To overcome these limitations, the authors propose a physics‑informed neural field (PINN‑based) framework that reconstructs the near‑surface broadband pressure field from a sparse set of pressure measurements and directly infers the normalized complex surface impedance ζ = Z/(ρ₀c₀).

Core Methodology
The approach hinges on three innovations: (1) a parallel multi‑frequency neural network architecture, (2) sinusoidal representation networks (SIREN) combined with modified multilayer perceptrons (ModMLP) for each frequency, and (3) a composite loss that enforces data fidelity, Helmholtz PDE residuals, spatial homogeneity of impedance, and frequency smoothness. The spatial domain is split into three coordinate sets: boundary points (S_b), volumetric points (S_v) where the Helmholtz residual is evaluated, and sampling points (S_s) where actual microphones record pressure. Two parallel layers of microphones are placed at distances d₁ (surface‑to‑first‑layer) and d₂ (inter‑layer spacing). By using automatic differentiation, the normal pressure gradient ∂p/∂n is obtained from the learned pressure field, eliminating the need for particle‑velocity sensors (p‑u probes) that are prone to vibration and temperature sensitivity.

Hardware Design
A compact microphone array is built with either 3 × 3 (9 channels) or 4 × 4 (16 channels) elements, each spaced 25 mm apart. The array’s low channel count drastically reduces hardware complexity compared with traditional NAH/ESM setups that require dozens to hundreds of microphones or mechanical scanning rigs.

Numerical Validation
Simulations under ideal plane‑wave incidence in an anechoic environment explore the sensitivity of the method to array geometry. Two benchmark surfaces are considered: a porous absorber (highly dissipative) and a near‑rigid plate (low dissipation). A parametric sweep over d₁ and d₂ reveals distinct behavior:

• Porous absorber: error surfaces are flat; both 3 × 3 and 4 × 4 arrays achieve low mean absolute errors (MAE) in absorption (α) and impedance (ζ) across a wide range of distances. The 4 × 4 array maintains MAE ≈ 0.01 for α and ≤ 0.04 for ζ even when d₁ = 20 mm, which is practical for many installations.
• Near‑rigid surface: error surfaces are sharply peaked. Accurate reconstruction requires the microphones to be very close to the surface (d₁ ≈ 5 mm, d₂ ≈ 10 mm) because the pressure gradient vanishes, reducing the signal‑to‑noise ratio for the inferred velocity. The 4 × 4 array outperforms the 3 × 3 array, but both are highly sensitive to placement.

Training convergence analysis shows that porous cases converge within 2,500–5,000 epochs (≈ 30–60 s on a consumer‑grade GPU), whereas the rigid case needs more epochs and exhibits larger fluctuations in the absorption error early in training. Importantly, low absorption error does not guarantee low impedance error, underscoring the non‑uniqueness of the α → ζ mapping.

Noise Robustness
Adding complex Gaussian noise with SNR ranging from 30 dB to 70 dB degrades performance predictably. The porous absorber remains robust (MAE < 0.05 even at 30 dB), while the rigid case deteriorates more rapidly, confirming the need for higher SNR or closer sensor placement for low‑impedance surfaces.

Experimental Validation
The authors validate the framework in two laboratory settings: an anechoic chamber and a reverberant room. The compact array records pressure fields generated by a single loudspeaker. The neural field is trained on the measured data, and the inferred impedance matches reference measurements (obtained with a calibrated impedance tube) within a few percent for the porous sample and within acceptable bounds for the rigid sample when the optimal geometry is used.

Vehicle Cabin Demonstration
A virtual vehicle cabin model, featuring complex geometry, multiple reflective surfaces, and modal interference, serves as a realistic testbed. Using only four microphones placed on a seat back, the method reconstructs the pressure field throughout the cabin and extracts the surface impedance of the seat material. The results align closely with ground‑truth data derived from a full‑scale measurement campaign, demonstrating that the approach scales to highly reverberant, non‑ideal environments.

Key Contributions

1. Parallel Multi‑Frequency Neural Field – Enables broadband inference without the optimization difficulties of traditional PINNs that treat frequency as an input variable.
2. Automatic Velocity Estimation – Removes reliance on particle‑velocity sensors, simplifying hardware and improving robustness to environmental perturbations.
3. Sparse Sensor Strategy – Shows that as few as nine pressure sensors can yield accurate impedance estimates for dissipative materials, while sixteen sensors suffice for low‑impedance surfaces.
4. Real‑World Validation – Demonstrates performance in both controlled laboratory conditions and a realistic automotive cabin, confirming applicability beyond idealized scenarios.

Future Directions
The authors suggest extending the framework to handle multiple simultaneous sources, non‑planar incident fields, and nonlinear material behavior. Incorporating Bayesian inference could provide uncertainty quantification for the estimated impedance, which is valuable for design optimization and risk assessment.

In summary, this work delivers a practical, physics‑driven machine‑learning pipeline that bridges the gap between laboratory acoustic characterization and real‑world in‑situ measurement, offering rapid, accurate surface impedance retrieval with minimal hardware overhead.