Dynamic Gaussian Graph Operator: Learning parametric partial differential equations in arbitrary discrete mechanics problems

Chu Wang, Jinhong Wu, Yanzhi Wang, Zhijian Zha, Qi ZhouEngineering Applications of AI 2025 · DOI

TL;DR

This paper introduces the Dynamic Gaussian Graph Operator (DGGO), a novel neural operator designed to solve parametric partial differential equations (PDEs) on arbitrarily discretized computational domains. DGGO overcomes the limitations of existing methods that struggle with non-uniform meshes, demonstrating superior accuracy and generalization capabilities across various mechanics problems and an engineering application.

What Problem Does This Paper Address?

Solving complex physical systems often relies on parametric partial differential equations (PDEs). While deep learning has shown promise in this area through operator learning, current neural operators are largely confined to constant and uniform discretizations of computational domains. This limitation hinders their ability to generalize to real-world scenarios where meshes are often irregular and non-uniform. Formulating and solving these PDEs numerically can also be computationally expensive and require extensive prior knowledge. This paper aims to bridge this gap by developing a neural operator that can effectively handle arbitrary discretizations, making it more applicable to a wider range of complex physical problems.

Methods

The core innovation is the Dynamic Gaussian Graph Operator (DGGO), which comprises three main components:

  • Forward DGG Kernel: This component maps input vectors from a general Euclidean space to a high-dimensional uniform metric space. It utilizes a dynamic message-passing graph structure with Gaussian kernel similarities to capture interrelations within the data. This process is designed to achieve spatial transformation and prepare the data for spectral analysis.
  • Spectral Transform-based Operator: A replaceable block that leverages spectral transforms (specifically, the Fourier Transform in this work, inspired by Fourier Neural Operator - FNO) to localize metric vectors in both spatial and frequency domains. This provides additional constraints on the solution space.
  • Inverse DGG Kernel: This component reconstructs the solution vectors back into the general Euclidean space from the transformed metric vectors.

The DGGO iteratively refines hidden state vectors through a sequence of layers, approximating integrals using Riemann sums and employing a dynamic message-passing graph to model local similarities. This approach allows for handling arbitrary discrete points and time-dependent inputs.

Datasets

The paper evaluates DGGO on several benchmark datasets for numerical problems, including:

  • 1D Burgers equation
  • 1D time-dependent wave advection equation
  • 2D Darcy flow equation
  • 2D time-dependent Navier-Stokes equation

These datasets involve various spatial and temporal resolutions, including both uniform and non-uniform discretizations. The authors also present an engineering application on hyper-elastic material, for which they generated training and testing datasets.

Key Insights and Ideas

  • Arbitrary Discretization is Key: The paper highlights that real-world problems rarely conform to uniform grids. DGGO's ability to handle arbitrary discretizations is a significant step towards practical applications of neural operators.
  • Spatial Transformation for Uniformity: The DGG kernel's core function is to transform non-uniform input spaces into a latent uniform metric space. This transformation is crucial for enabling spectral methods to work effectively.
  • Graph-based Message Passing for Local Interactions: Using dynamic message passing with Gaussian kernels allows DGGO to effectively capture local relationships and similarities between discrete points, even in irregular domains.
  • Synergy of Spectral and Graph Methods: DGGO combines the strengths of graph-based learning for spatial transformation with spectral methods for frequency localization, creating a powerful framework for PDE solving.
  • Spatial Constraint Matters: Ablation studies show that including spatial coordinate residuals in the loss function significantly improves accuracy, underscoring the importance of spatial information in the learning process.

Key Results

  • Superior Accuracy: DGGO consistently outperforms state-of-the-art neural operators (FNO, WNO, DeepONet, POD-DeepONet) and traditional deep neural networks (U-Net, PointNet) across various numerical cases, often by a significant margin. For instance, in the 1D Burgers equation, DGGO achieved errors as low as 2.80% compared to FNO's 6.56% at a resolution of 512.
  • Robustness to Non-uniformity: The method demonstrates strong performance on non-uniform discrete schemes, a significant advantage over existing spectral transform-based operators.
  • Engineering Application Success: In the hyper-elastic material stress forecasting task, DGGO achieved a low relative error of 1.52% for unseen void shapes, showcasing its applicability to complex engineering problems.
  • Efficiency: While not explicitly quantified in terms of speedup, the method's ability to generalize to new discretizations without retraining implies significant efficiency gains for practical applications.

Limitations and Future Work

  • Hyperparameter Tuning: The paper mentions that the DGG kernel introduces hyperparameters (number of discrete nodes, Gaussian kernel variance, metric space dimension) that need to be selected based on the specific problem.
  • Computational Cost: While not a primary focus, the iterative nature of the DGG kernel and the use of message passing might imply a higher computational cost compared to simpler architectures for very large-scale problems.
  • Future Work: The authors suggest that the spectral transform block can be replaced with other wavelet-based transforms, indicating potential for further customization and optimization. The success in engineering applications also opens doors for exploring more complex physical systems.

Summary updated

BibTeX
@article{wang2025dynamic,
  author = {Chu Wang and Jinhong Wu and Yanzhi Wang and Zhijian Zha and Qi Zhou},
  title = {{Dynamic Gaussian Graph Operator: Learning parametric partial differential equations in arbitrary discrete mechanics problems}},
  journal = {Engineering Applications of AI},
  year = {2025},
  doi = {10.1016/j.engappai.2025.110405},
}
Figure