Acoustics-Informed Neural Networks

• Acoustics-Informed Neural Networks (AINNs) are neural models that integrate fundamental acoustic physics, including the wave and Helmholtz equations, to ensure physically consistent predictions.
• AINNs employ customized architectures and PDE-informed losses to accurately reconstruct acoustic fields, enabling applications such as room impulse response mapping and parameter identification.
• By fusing data-driven learning with first-principles constraints, AINNs achieve robust and efficient performance even in scenarios with sparse, undersampled, or noisy measurements.

Acoustics-Informed Neural Networks (AINNs) refer to a class of neural network models that are explicitly constrained or guided by the fundamental physics of sound propagation, including domain-specific PDEs such as the wave or Helmholtz equations, and/or embedded structured priors reflecting acoustical scene properties. AINNs bridge the gap between conventional data-driven learning and first-principles modeling, enabling physically consistent, robust, and efficient acoustic field inference, surrogate modeling, parameter identification, and scene reconstruction—even when measurements are incomplete, undersampled, or corrupted by noise.

1. Mathematical Foundations and Key Principles

AINNs leverage the governing equations of acoustics, such as the scalar wave equation for pressure fields ($\nabla^2 h - \frac{1}{c^2} \partial_t^2 h = 0$), the Helmholtz equation for time-harmonic fields ($\nabla^2 P + k^2 P = 0$), or coupled momentum/continuity systems for direction-aware or multi-field models. The central methodological advance is to embed these equations as “physics-informed” constraints on the network’s output, typically as a “soft” (loss term) or “hard” (architectural) constraint.

• Data-fitting and PDE-informed losses: AINNs optimize a composite objective $$\mathcal{L} = \mathcal{L}_{\mathrm{data}} + \lambda \mathcal{L}_{\mathrm{PDE}}$$, where $\mathcal{L}_{\mathrm{PDE}}$ penalizes pointwise residuals of the relevant PDE(s)—e.g., $$|| \frac{1}{c^2} \partial_t^2 h_{\theta} - \nabla_r^2 h_{\theta} ||^2$$—and $\mathcal{L}_{\mathrm{data}}$ enforces agreement with observed microphone data or impulse responses (Pezzoli et al., 2023, Ma et al., 2024, Schoder et al., 18 May 2025).
• Operator regularization: Some architectures, e.g., for direction-aware Ambisonic fields, add constraints reflecting not just the wave equation but also the linearized momentum $\nabla p + \rho_0 \partial_t u = 0$ and continuity $$\rho_0 \nabla \cdot u + \frac{1}{c_0^2} \partial_t p = 0$$ equations, yielding significantly stronger inductive biases for multi-component fields (Masuyama et al., 9 Jul 2025).
• Physical parameter inference: Inverse AINN variants not only reconstruct fields but also infer physical parameters (e.g., loss coefficients, radiation impedances) as part of the optimization—trained either as free parameters or with Bayesian/posterior approaches (Luan et al., 18 May 2025, Yokota et al., 2024, Huang et al., 2023).

2. Neural Architectures and Implementation Strategies

AINNs employ diverse neural architectures but typically favor compact, low-depth multi-layer perceptrons (MLPs) with carefully chosen activations to enable accurate representation (including high-frequency modes) and efficient computation of derivatives via automatic differentiation.

Architecture	Purpose	Representative Papers
SIREN MLPs (sinusoidal)	High-frequency field representation, PINNs	(Pezzoli et al., 2023, Masuyama et al., 9 Jul 2025)
Standard FNNs	Surrogate models, field parameterization	(Borrel-Jensen et al., 2021, Ma et al., 2024)
Galerkin/Hard-constrained	Embed solution eigenspaces, exact BCs	(Ozan et al., 2023)
Hierarchical/Residually connected	Memory, stability, efficiency	(Kalthoff et al., 26 Nov 2025, Yokota et al., 2024)
Explicit context and geometric features	Fuse mesh-based scene data	(Si et al., 18 Sep 2025)

Key techniques include:

• Periodic/sinusoidal activations (SIREN, Snake) to avoid spectral bias and enable smooth computation of derivatives, essential for high-frequency acoustic phenomena (Pezzoli et al., 2023, Masuyama et al., 9 Jul 2025, Luan et al., 18 May 2025).
• Decoupled real/imaginary modeling for time-harmonic fields, leveraging separate compact MLPs for Re/Im parts, proven superior to unified architectures in multiple SFR tasks (Ma et al., 2024).
• Architectural hard constraints embedding the eigenspace or boundary conditions (Galerkin NNs), enabling physically constrained extrapolation with minimal data (Ozan et al., 2023).
• Physics-guided latent variables (e.g., local grid latents, proximity/raycast features from meshes) as context channels, directly coupling geometric environment information to field prediction (Si et al., 18 Sep 2025, Luo et al., 2022).

3. Representative Applications

AINNs have been deployed in a broad spectrum of acoustic signal processing problems, exemplified below.

Room Impulse Response and Sound Field Reconstruction

• Sparse-to-dense RIR mapping: By leveraging the acoustic wave equation, AINNs reconstruct full multichannel RIRs across arrays from sparse microphone observations, outperforming compressive sensing and standard deep priors, and delivering robust performance with lightweight architectures (Pezzoli et al., 2023).
• Sound field mapping in bounded regions: Compact AINNs regularized by the Helmholtz equation outperform classical cylinder harmonics and SVD-based decompositions in both pressure and gradient field reconstruction, achieving strong robustness to noise and data scarcity (Ma et al., 2024).
• High-dimensional FOA modeling: Direction-aware AINNs for first-order Ambisonic RIRs enforce both wave and momentum/continuity constraints, producing physically consistent spatial and directional field reconstructions (Masuyama et al., 9 Jul 2025).

Parameter Estimation and Inverse Problems

• Loss and radiation parameter identification: PINN-based AINNs identify viscothermal losses and unknown boundary parameters (e.g., $\alpha, \beta$ in radiation models), demonstrating strong noise-tolerance and outperformance over classical TOM under uncertain conditions (Luan et al., 18 May 2025, Yokota et al., 2024).
• Bayesian and neural field regression: Neural-physical AINNs coupled with Bayesian inference yield robust, UQ-calibrated estimates of propagation coefficients and room parameters directly from sparse frequency-domain data (Huang et al., 2023).

Geometry-Infused Scene-Aware Acoustic Prediction

• Mesh-infused neural acoustic fields: By extracting explicit geometric features (rays, normals, occupancy histograms) from rough or noisy meshes and fusing these into neural RIR predictors, AINNs achieve marked improvements in T₆₀, C₅₀, and EDT, especially in low-data regimes and under mesh noise. Ablation confirms the critical value of explicit geometry cues (Si et al., 18 Sep 2025).

Acoustic Hardware Implementation and Digital Twin Training

• Physical realizability: AINNs trained under passive constraints (non-negative weights, no bias, intensity-based activations) map directly onto physical waveguide hardware (SincHSRNN, HSRNN), enabling low-power, analog, wave-based neural computation with learnable bandpass filters and intensity-based logic operations (Kalthoff et al., 26 Nov 2025).

Beamforming and Virtual Microphone Enhancement

• Microphone array extension: AINNs trained to map boundary-point measurements to virtual microphone predictions eliminate deep nulls and spatial aliasing in circular array beamforming, surpassing solutions requiring more physical microphones and matching concentric array performance (Zhao et al., 2024).

4. Performance Metrics and Quantitative Evaluation

AINN performance is conventionally evaluated with error metrics sensitive to both physical plausibility and acoustic signal fidelity:

• Normalized mean square error (NMSE) in dB: Assesses field or RIR reconstruction accuracy, typically over test positions (Pezzoli et al., 2023, Ma et al., 2024).
• Spectral MSE, T₆₀ error, C₅₀ and EDT: Standard in neural-acoustic field benchmarks, capturing reverberation and early/late energy ratios (Luo et al., 2022, Si et al., 18 Sep 2025).
• Correlation score (PCC) and cross-correlation error: Used for FOA and binaural fields, reporting spatial/temporal coherence (Masuyama et al., 9 Jul 2025).
• Model efficiency: Number of parameters (typically $\sim$10^5–10^6), inference runtime per query (up to real-time), and required sampling density (e.g., $\nabla^2 P + k^2 P = 0$0 collocation points/wavelength) (Pezzoli et al., 2023, Schoder et al., 18 May 2025).
• Comparative baselines: AINNs generally outperform compressive sensing, standard DNNs/U-Nets, SVD/basis decompositions, and sometimes physical model interpolations, especially under sparse, noisy, or out-of-domain conditions (Ma et al., 2024, Zhao et al., 2024, Si et al., 18 Sep 2025).

5. Design, Training, and Practical Considerations

AINNs require careful coordination of physics induction and machine learning principles:

• Sampling: Uniform, randomized, or geometry-aware sampling over spatial/temporal domains to ensure PDE and boundary terms are sufficiently actuated.
• Loss balancing: Manual or adaptive weighting of PDE, BC, and data terms to keep gradients stable and optimization tractable (Pezzoli et al., 2023, Masuyama et al., 9 Jul 2025).
• Activations: Sine or periodic activations for smooth, non-spectrally biased outputs, and to support higher derivatives for stiff PDEs (Pezzoli et al., 2023, Schoder et al., 18 May 2025).
• Training: Adam or similar optimizers, large-batch/“full-batch” settings (especially for physics collocation), auto-differentiation for derivatives, with convergence typically in thousands to hundreds of thousands of steps (Pezzoli et al., 2023, Schoder et al., 18 May 2025).
• Pretraining and transfer: For high-frequency or large domains, pretraining on Green’s functions or analytic solutions, freezing early layers, or incremental learning (“discrepancy learning”) can accelerate convergence (Schoder et al., 18 May 2025).
• Robustness: Regularization (row-clustering, L1/L2 on weights), ensemble training, and explicit context fusion boost out-of-sample and noise robustness (Huang et al., 2023, Si et al., 18 Sep 2025).

6. Extensions and Limitations

AINNs have been realized for 1D, 2D, and 3D domains, including tubes, rooms, and free-space with various BCs (Neumann, Dirichlet, impedance), but have several limitations:

• Computational cost: Training remains significantly slower than a single numerical FDTD/FEM sweep but inference is highly efficient and storage is dramatically reduced (Borrel-Jensen et al., 2021, Ma et al., 2024).
• Physics regime: Most results to date focus on the early or low-frequency part of the RIR, linear and stationary fields, and idealized materials. Nonlinear, nonstationary, and highly heterogeneous domains present open challenges (Luan et al., 18 May 2025, Masuyama et al., 9 Jul 2025).
• Order and dimensionality: Scaling to high-frequency, high-dimensional or high-order Ambisonic fields requires deeper architectures, more complex priors, and/or hierarchical/multiscale schemes (Schoder et al., 18 May 2025, Masuyama et al., 9 Jul 2025).
• Parameter identifiability: For some loss parameters (e.g., viscous tube losses), the data may not be sufficiently informative for accurate estimation without supplementary measurements (Yokota et al., 2024).

7. Research Impact and Outlook

AINNs represent a convergence of classical acoustics, neural representation theory, and domain-knowledge-infused learning. They provide a general, extensible framework for acoustic field prediction, system identification, and real-time signal processing under realistic constraints and with limited data. Ongoing research seeks to expand their applicability to:

• Full-length RIRs and late reverberation via hybrid time-frequency PINNs
• Active hardware inference systems with direct mapping from digital-twin AINNs to physically instantiated acoustic processors
• Enhanced cross-modal (audio-visual-spatial) learning utilizing physically informed acoustic latents and explicit geometry coupling
• Scalable domain decomposition, multi-GPU parallelism, and real-world measurement generalization (Masuyama et al., 9 Jul 2025, Ma et al., 2024, Kalthoff et al., 26 Nov 2025).

AINNs are a foundational development in modern physical machine learning for acoustics and scene-aware signal processing, combining efficiency, interpretability, and physically valid generalization across a range of scenarios and hardware modalities (Pezzoli et al., 2023, Masuyama et al., 9 Jul 2025, Kalthoff et al., 26 Nov 2025, Ma et al., 2024, Si et al., 18 Sep 2025).