Fourier Heuristic Pinns to Solve the Biharmonic Equations Based on its Coupled Scheme

Xian Li, Yujia Huang, Jinran WuSSRN 2024 · DOI

TL;DR

This paper introduces FCPINN, a novel deep learning approach that combines Fourier feature mapping with a coupled scheme to efficiently solve high-order biharmonic equations. FCPINN demonstrates superior accuracy and speed compared to existing methods, particularly for complex and high-dimensional problems.

What Problem Does This Paper Address?

Solving high-order partial differential equations (PDEs), such as the biharmonic equation, is crucial in many scientific and engineering fields. Traditional numerical methods often struggle with computational complexity, especially for problems with intricate boundary conditions or in high-dimensional spaces, leading to the "curse of dimensionality." While Physics-Informed Neural Networks (PINNs) offer a mesh-free alternative, they can be computationally demanding for high-order derivatives and may suffer from spectral bias, limiting their ability to capture high-frequency information. This paper aims to overcome these limitations by developing a more efficient and accurate PINN-based method.

Methods

The core of the FCPINN approach lies in two key innovations:

  • Coupled Scheme: The fourth-order biharmonic equation is decomposed into two second-order Poisson equations by introducing an auxiliary variable. This reduces the need to compute high-order derivatives directly, simplifying the problem and improving computational efficiency.
  • Fourier Feature Mapping: To combat the spectral bias inherent in standard neural networks, FCPINN incorporates Fourier feature mapping in the first hidden layer. This technique, using sine and cosine functions, allows the network to effectively learn and represent high-frequency components of the solution, mimicking Fourier expansion.

The FCPINN architecture uses a single multi-output neural network to approximate both the solution and the auxiliary variable. The training process involves minimizing a combined loss function that includes residuals from both the governing equations and the boundary conditions.

Datasets

Not discussed in the paper. The paper focuses on numerical experiments using synthetic data generated from known analytical solutions for various biharmonic equation problems.

Key Insights and Ideas

  • Decomposition is Key for High-Order PDEs: Breaking down the high-order biharmonic equation into simpler, coupled second-order Poisson equations significantly reduces computational complexity and the need for high-order derivatives, a common bottleneck in PINNs.
  • Fourier Features Conquer Spectral Bias: Applying Fourier feature mapping in the initial layers of the neural network is crucial for enabling the model to effectively capture high-frequency components of the solution, which is often a challenge for standard deep learning models.
  • Synergy of Coupled Schemes and Fourier Features: The combination of a coupled scheme and Fourier feature mapping in FCPINN leads to a synergistic effect, resulting in a method that is both computationally efficient and highly accurate, outperforming existing approaches.
  • Multi-Output Networks Simplify Architecture: Using a single multi-output network for both the solution and auxiliary variables streamlines the architecture and potentially reduces computational overhead compared to using separate networks.

Key Results

The FCPINN model consistently demonstrated superior performance across various biharmonic equation problems, including linear and nonlinear cases, and in different dimensional spaces (2D, 3D, and 8D).

  • Accuracy: FCPINN achieved significantly lower relative errors (REL) compared to standard PINN and MIM (Mixed Residual Method) approaches. For instance, in Example 4.1 (Dirichlet boundary), FCPINN achieved an REL of $4.905 \times 10^{-5}$, while MIM had $2.276 \times 10^{-4}$ and PINN had $3.342 \times 10^{-4}$.
  • Speed: While FCPINN's computation time was comparable to MIM and slightly slower than MIM in some cases, it was considerably faster than standard PINN. For Example 4.1, FCPINN took 981.735 seconds, MIM took 792.726 seconds, and PINN took 1938.947 seconds.
  • High-Dimensional Performance: In 8D problems, FCPINN maintained its accuracy advantage over MIM, with a computation speed close to MIM, highlighting its effectiveness in high-dimensional settings.
  • Robustness: The method showed robustness across different boundary conditions (Dirichlet and Navier) and problem complexities.

Limitations and Future Work

  • Computational Resource Demands for High Dimensions: While FCPINN performs well in high dimensions, the paper notes that for very high-dimensional problems (e.g., 8D), the computational resource requirements can still be significant, limiting the scale of experiments that can be conducted.
  • Future Work: The authors suggest further expanding and refining the neural network model to tackle other higher-order PDEs. They also imply that exploring different activation functions and network architectures could lead to further improvements.

Summary updated

BibTeX
@misc{li2024fourier,
  author = {Xian Li and Yujia Huang and Jinran Wu},
  title = {{Fourier Heuristic Pinns to Solve the Biharmonic Equations Based on its Coupled Scheme}},
  year = {2024},
  doi = {10.2139/ssrn.5014416},
  howpublished = {SSRN},
}
Figure