Solving the Nonlinear Westervelt Equation in Ultrasound Propagation Using Physics-Informed Neural Networks

Luke Oluwaseye JoelarXiv 2026 · arXiv

This paper presents a novel approach to solving the Westervelt equation—the mathematical backbone of nonlinear acoustics—using Physics-Informed Neural Networks (PINNs). By embedding physical laws directly into the neural network's learning process, the researcher successfully models how high-intensity ultrasound waves distort and decay as they travel through biological tissue, offering a mesh-free alternative to traditional numerical solvers.

TL;DR

The study applies PINNs to solve the 1D nonlinear Westervelt equation across three distinct physical regimes: linear, shock-dominated, and dissipation-dominated. By using a sinusoidal activation function, the model accurately captures complex wave phenomena like harmonic generation and amplitude decay without the need for traditional computational meshes.

What Problem Does This Paper Address?

Accurately modeling ultrasound propagation is vital for both diagnostic imaging and therapeutic applications like High-Intensity Focused Ultrasound (HIFU). However, biological tissues are "thermoviscous fluids," meaning they are both nonlinear and attenuating.

The Westervelt equation describes this behavior, but it is a challenging nonlinear partial differential equation (PDE). Traditional numerical methods (like Finite Difference or Finite Element methods) often struggle with this equation because:

  1. Nonlinearity causes waveforms to steepen into shocks, requiring extremely fine (and computationally expensive) meshes.
  2. Attenuation introduces higher-order derivatives that can lead to numerical instability.
  3. Oscillatory behavior is difficult for standard neural networks to represent.

The author seeks to fill the gap where traditional solvers fail by leveraging the flexibility of deep learning constrained by physics.

Methods

The core methodology relies on Physics-Informed Neural Networks (PINNs). Unlike standard neural networks that learn purely from data, PINNs use the PDE itself as a supervisor.

  • The Architecture: A fully connected network with 3 hidden layers and 64 neurons per layer.
  • Sinusoidal Activation: Instead of the common ReLU activation, the author uses a sin function. This is a critical choice for wave physics, as it naturally represents the periodic, oscillatory nature of acoustic pressure.
  • The Loss Function: The network is trained to minimize a composite loss function consisting of four parts:
  • PDE Residual: How well the predicted pressure satisfies the Westervelt equation.
  • Initial Conditions: Ensuring the medium starts at rest.
  • Source Boundary: Matching the input from a Gaussian-enveloped cosine pulse (simulating a real ultrasound transducer).
  • Absorbing Boundary: Implementing a Sommerfeld-type condition at the far end of the domain to prevent unphysical reflections.
  • Optimization: The model was trained using the Adam optimizer for 30,000 epochs with a learning rate of $10^{-3}$.

Datasets

This study does not use external experimental datasets. Instead, it evaluates the PINN framework against three synthetic cases designed to represent different tissue properties found in medical ultrasound:

  1. Case 1 (Weak Nonlinearity, Low Attenuation): Mimics water-like media or homogeneous soft tissue at low pressures.
  2. Case 2 (Stronger Nonlinearity, Low Attenuation): Represents typical soft tissue where wave distortion is significant.
  3. Case 3 (Strong Nonlinearity, High Attenuation): Represents dense or highly attenuating tissues often targeted in therapeutic ultrasound.

Key Insights and Ideas

  • Parameter Separation: The study confirms that the two primary parameters, $\eta$ (nonlinearity) and $\gamma$ (attenuation), have distinct roles. $\gamma$ scales the rate of amplitude decay (energy absorption), while $\eta$ governs frequency-dependent phase speeds and harmonic generation.
  • Energy Feeding: In Case 2 (high nonlinearity), the author observed a "peak creep" where the amplitude briefly rises above the source amplitude before decaying. This reflects energy being fed into the primary waveform from the nonlinear interaction as it propagates.
  • Harmonic Plateaus: The frequency analysis revealed that in nonlinear cases, the "deep target" spectrum develops a broad plateau at higher frequencies. This indicates that the PINN is successfully capturing the transfer of energy from the fundamental frequency to higher harmonics—a hallmark of nonlinear acoustics.

Key Results

The PINN framework demonstrated high accuracy and convergence across all test cases:

  • Convergence: The PDE residual converged to at least $10^{-6}$, and the total PINN loss reached $10^{-5}$ within 30,000 epochs.
  • Waveform Fidelity: As seen in the figure below, the PINN accurately tracks the pulse as it moves through the tissue domain.
Figure 1: Case 1: Loss components and acoustic propagation field.
  • Nonlinear Capture: In Case 2, the model captured the development of a "residual tail" after each wave trough, a subtle physical detail that traditional low-resolution solvers often miss.
  • Mesh-Free Advantage: The solution is continuous across the entire spatio-temporal domain, allowing for pressure predictions at any coordinate $(x, t)$ without interpolation between grid points.

Limitations and Future Work

While promising, the author acknowledges several limitations:

  • 1D Geometry: The current work is limited to one dimension. Clinical ultrasound involves complex 3D beam patterns and focusing, which will require scaling the PINN architecture.
  • Homogeneous Media: The study assumes uniform tissue properties. Real biological environments are heterogeneous (e.g., waves passing through fat, then muscle, then bone), requiring spatially varying parameters.
  • Normalization: The parameters used were dimensionless. Future work will involve mapping these directly to physical tissue measurements (e.g., specific $B/A$ nonlinearity ratios and absorption coefficients in dB/cm/MHz).

Summary updated

Abstract

Modeling the nonlinear acoustic wave propagation in biological tissues needs accuracy, especially in ultrasound imaging and therapeutic applications. The 1D Westervelt equation, \( p_{xx} - p_{tt} + γ\, p_{ttt} + η\,(p^{2})_{tt}=0, \) could be used to describe a finite-amplitude ultrasound propagation in these tissues. This equation accounts for nonlinearity, \(η\), and attenuation, \(γ\). To solve this equation, we observe that traditional solvers may struggle to do so correctly. This is due to the difficulties arising from the nonlinear nature of the equation. We solve the equation by defining three synthetic cases - linear (\(η=0.001\), \(γ=10^{-6}\)), shock-dominated (\(η=0.02\), \(γ=10^{-6}\)), and dissipation-dominated (\(η=0.01\), \(γ=5\times10^{-4}\)) - and using Physics-Informed Neural Networks (PINNs) to showcase the solution. A sinusoidal activation function is used to capture the oscillatory behavior of the problem. The results of our study, with PDE residual converging to at least \(10^{-6}\), boundary condition loss to \(10^{-5}\), and the PINN loss converging to \(10^{-5}\) for each case at \(30,000\) epochs, show that \(γ\) scales the rate of amplitude decay, while \(η\) governs frequency-dependent phase speeds.
BibTeX
@misc{joel2026solving,
  author = {Luke Oluwaseye Joel},
  title = {{Solving the Nonlinear Westervelt Equation in Ultrasound Propagation Using Physics-Informed Neural Networks}},
  year = {2026},
  eprint = {2611.05130},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.05130},
}
Figure