BubbleONet: A Physics-Informed Neural Operator for High-Frequency Bubble Dynamics

Yunhao Zhang, Lin Cheng, Aswin Gnanaskandan, Ameya D. JagtaparXiv 2025 · DOI · arXiv

Introduction

Microscopic gas bubbles play a critical role in diverse applications, ranging from ultrasound-mediated drug delivery to industrial jet cleaning. Modeling their behavior, however, presents significant computational challenges. At high frequencies, bubbles undergo rapid oscillations and exhibit nonlinear instabilities governed by "stiff" ordinary differential equations (ODEs) such as the Rayleigh–Plesset and Keller–Miksis equations.

While traditional numerical solvers provide high accuracy, they are computationally intensive, particularly when simulating systems with millions of interacting bubbles. Deep learning offers a potential path toward accelerated simulation, but standard neural networks often suffer from spectral bias—a tendency to prioritize smooth, low-frequency patterns while failing to capture the high-frequency "wiggles" that are physically essential in bubble dynamics.

Methodology: The BubbleOKAN Architecture

To address these challenges, researchers introduced BubbleOKAN, a Kolmogorov-Arnold Network-based Deep Operator Network. This architecture integrates three core principles to achieve high-frequency resolution and physical consistency.

The BubbleOKAN architecture consists of an Operator Layer with Branch and Trunk networks, integrated with a physics-informed Operation Layer.
  1. DeepONet Framework: The model employs a dual-network structure. A "branch" network processes input pressure profiles, while a "trunk" network handles coordinates such as time (t) and initial radius (R0).
  2. Kolmogorov-Arnold Networks (KANs): Unlike traditional Multi-Layer Perceptrons (MLPs) with fixed activation functions, KANs utilize learnable basis functions on the edges. This enhances interpretability and parametric efficiency.
  3. Physics-Informed Loss: The training process minimizes a loss function that includes the residual of the governing physical equations (L_ode), ensuring that predictions adhere to the laws of fluid mechanics.

Two-Step Training and Spline Basis

A critical innovation in BubbleOKAN is its two-step training strategy designed to overcome spectral bias. First, the trunk network is trained to establish a robust functional basis using spline basis functions. The authors found that splines significantly outperform Radial Basis Functions (RBFs) in resolving high-frequency oscillations. Second, the branch network is trained using the fixed basis from the trunk. This is complemented by a continual learning strategy, where the model is exposed to data in frequency stages (High -> Medium -> Low) to prevent the loss of high-frequency details.

Experimental Datasets

The model was evaluated using synthetic datasets generated by the APECSS numerical solver:

  • Rayleigh–Plesset (R-P) Equation: 3,000 datasets with a fixed initial radius (50 um), and varying pressure amplitudes (1-10 x 10^5 Pa) and frequencies (200-2000 kHz).
  • Keller–Miksis (K-M) Equation: Datasets accounting for liquid compressibility.
  • Multi-Radii K-M: A more complex set involving five initial radii (50, 60, 70, 80, 90 um) to test generalization across bubble sizes.

Results and Discussion

BubbleOKAN demonstrated superior accuracy compared to vanilla DeepONet and other state-of-the-art operators like Fourier Neural Operators (FNO), Wavelet Neural Operators (WNO), and Convolutional Neural Operators (CNO).

Comparison of bubble radius predictions and absolute errors at 676 kHz (a) and 1654 kHz (b).

In high-frequency test cases (e.g., 2000 kHz), the L2-norm error was reduced from 0.287% in standard models to as low as 0.020%. As shown in the figure above, the model accurately captures the fine-grained oscillations across different frequency regimes, maintaining low absolute error throughout the temporal domain.

The model also showed strong generalization capabilities for "out-of-distribution" pressure amplitudes. However, it encountered difficulties with frequency extrapolation and resolving dynamics near the resonance frequency, where physical sensitivity is extreme.

Conclusion

BubbleOKAN represents a significant step forward in applying neural operators to stiff, high-frequency physical systems. By combining the structural advantages of DeepONets with the flexibility of KANs and spline-based basis functions, the model effectively mitigates spectral bias. Future work will focus on bridging the "resonance gap" and extending the model's temporal range to handle more complex turbulent flow scenarios.

Summary updated

BibTeX
@misc{zhang2025bubbleonet,
  author = {Yunhao Zhang and Lin Cheng and Aswin Gnanaskandan and Ameya D. Jagtap},
  title = {{BubbleONet: A Physics-Informed Neural Operator for High-Frequency Bubble Dynamics}},
  year = {2025},
  eprint = {2508.03965},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.03965},
}
Figure