LFR-PINO: A Layered Fourier Reduced Physics-Informed Neural Operator for Parametric PDEs

Jing Wang, Biao Chen, Hairun Xie, Rui Wang, Yifan Xia, Jifa Zhang, Hui XuarXiv 2025 · DOI · arXiv

Solving Partial Differential Equations (PDEs) is fundamental to engineering and scientific simulation, from modeling airflow over a wing to predicting heat dissipation in microchips. However, when these equations are "parametric"—meaning coefficients, boundary conditions, or initial states change—traditional numerical solvers become computationally prohibitive. While Physics-Informed Neural Networks (PINNs) offer a mesh-free alternative, they often require retraining for each new parameter set. Neural operators like FNO attempt to learn the mapping between parameter spaces but typically require vast amounts of labeled data.

Researchers from Shanghai Jiao Tong University and Zhejiang University have introduced LFR-PINO (Layered Fourier Reduced Physics-Informed Neural Operator), a framework designed to solve parametric PDEs with high accuracy and low memory overhead. By employing layered hypernetworks and frequency-domain weight reduction, LFR-PINO overcomes the "curse of dimensionality" prevalent in existing hypernetwork-based solvers.

The LFR-PINO Architecture

The core innovation of LFR-PINO lies in its approach to generating the weights of a solver network. Standard Hyper-PINNs use a single massive network to predict all weights, leading to exploding memory requirements. LFR-PINO introduces two key structural improvements to mitigate this:

  1. Layered Hypernetworks: Instead of one giant hypernetwork, LFR-PINO utilizes a dedicated, smaller hypernetwork for each individual layer of the primary solver network. This "divide and conquer" approach allows the model to specialize and simplifies the optimization landscape.
  2. Frequency-Domain Reduction: This is the framework's "secret sauce." Drawing on the "frequency principle" (the observation that neural networks often learn low-frequency patterns first), the authors project the network weights into the Fourier domain. Instead of predicting every individual weight, the hypernetworks only predict a small set of low-frequency Fourier coefficients. These coefficients are then transformed back into full weight matrices using an Inverse Fast Fourier Transform (IFFT).
The LFR-PINO framework architecture, illustrating the layered hypernetworks and the frequency-domain reduction process for weight generation.

Key Insights and Methodology

By focusing only on the most significant spectral features, the model captures the essential "physics" of the weights while ignoring high-frequency noise and redundant data. Several key insights drive this success:

  • Weight Continuity: The authors demonstrate that small changes in PDE parameters lead to smooth, continuous changes in the network weights. This continuity makes Fourier-domain compression highly effective—you don't need high-frequency components to describe a smooth transition.
  • Spectral Bias as a Feature: By explicitly forcing the hypernetwork to operate in the low-frequency regime, LFR-PINO inherently avoids the "noise" that often causes training instability in traditional Hyper-PINNs. This acts as a natural regularizer.
  • Universal to Specific Strategy: The framework uses a two-phase strategy. It first "pre-trains" a universal solver across many PDE variations and then allows for optional "fine-tuning" on a specific new equation, providing a balance between speed and precision.

Experimental Evaluation and Results

The researchers evaluated LFR-PINO on four classic PDE benchmarks: the Anti-derivative equation, the Advection equation, the Burgers equation, and a complex Diffusion-Reaction system.

LFR-PINO consistently outperformed baselines such as PI-DeepONet and standard Hyper-PINNs. Key results include:

  • Accuracy: In the pre-training stage, it achieved an error reduction of 22.8% to 68.7% across the test cases.
  • Efficiency: Thanks to the Fourier reduction, LFR-PINO reduced memory usage by 28.6% to 69.3% compared to standard Hyper-PINNs.
  • Stability: Theoretical analysis proved that low-frequency gradients dominate the training process in this architecture, leading to smoother convergence and faster fine-tuning.

Conclusion and Future Directions

LFR-PINO represents a significant step toward universal, efficient PDE solvers. While the layered architecture can require more training epochs, the gains in memory efficiency and accuracy are substantial. Future work will focus on extending this mesh-free approach to irregular, real-world geometries and combining spectral reduction with adaptive loss functions to tackle multi-physics coupling problems.

Summary updated

BibTeX
@misc{wang2025lfr,
  author = {Jing Wang and Biao Chen and Hairun Xie and Rui Wang and Yifan Xia and Jifa Zhang and Hui Xu},
  title = {{LFR-PINO: A Layered Fourier Reduced Physics-Informed Neural Operator for Parametric PDEs}},
  year = {2025},
  eprint = {2506.17582},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.17582},
}
Figure