Tuning and Optimizing the Finite Element Analysis with Elements of Large Nodal DOF on a Linux Cluster

Ji Wang, Lihong V. Wang, Qiang Sun, Rongxing Wu, Bin Huang, Jianke Du, Wei XiangarXiv 2015 · DOI · arXiv

Abstract

The finite element analysis (FEA) of high-frequency vibrations in quartz crystal plates is a critical step in designing precision resonators for filters and sensors. These devices utilize anisotropic materials and operate in the radiofrequency range, resulting in complex vibration spectra with numerous coupled modes. To achieve high precision, traditional 3D models require extremely fine meshes, leading to prohibitive computational costs. This paper reviews the development and optimization of FEMXR, a specialized parallel software suite designed to handle high-degree-of-freedom (DOF) problems using high-order plate theories on Linux clusters.

Introduction

Quartz crystal resonators are fundamental components in modern electronics, providing stable timing and frequency control. However, modeling their behavior is a significant challenge. The vibrations of interest—specifically thickness-shear modes—require representing sharp deformations across thin plates. While 3D analysis is the gold standard for accuracy, the resulting number of elements often exceeds practical limits. By utilizing the generalized Mindlin plate theory, researchers can reduce 3D piezoelectricity equations into 2D plate equations, significantly improving efficiency while maintaining 3D-level accuracy.

Software Architecture and Implementation

The core of this research is the FEMXR software, a parallel FEA suite implemented using a mix of C and Fortran. The architecture is designed to manage the unique challenges of high-order elements, which in this study involve up to 16 variables per node. This high nodal DOF density creates matrices with unusually large bandwidths compared to conventional structural FEA.

The architecture of the Parallel FEMXR software, illustrating the modular design and language integration across initialization, matrix generation, and solving stages.

As shown in the architecture diagram, the system is modular. Environment initialization and global matrix operations leverage MPI for parallelization on Linux clusters. The software distinguishes between free vibration (eigensolver) and forced vibration (linear equations solver) paths to address different engineering requirements.

Numerical Stack and Library Integration

Rather than reinventing fundamental numerical algorithms, FEMXR achieves high performance by integrating and tuning established high-performance computing (HPC) libraries. Solving the massive eigenvalue problems associated with million-plus DOF simulations requires a robust numerical stack.

The integration of specialized numerical libraries within the FEMXR eigensolver, highlighting the use of PARPACK, SuperLU_DIST, Aztec, and ParMETIS.

The eigensolver utilizes PARPACK for the Lanczos algorithm to extract eigenvalues and SuperLU_DIST for parallel sparse linear system solving. ParMETIS is employed for efficient mesh partitioning, while Aztec provides necessary preconditioning. The primary challenge identified by the authors is the "tuning and alignment" of these diverse libraries to work harmoniously on specific cluster architectures.

Key Insights and Results

  • Dimensionality Reduction: The transition from 3D to 2D Mindlin theory successfully reduces the total element count, though it increases the complexity of each individual element and the resulting matrix bandwidth.
  • Scalability: The software successfully processes simulations in the million-DOF range, a necessity for realistic resonator design.
  • Multi-Physics Capability: FEMXR effectively handles the coupling of mechanical vibrations and electrical fields, allowing engineers to calculate vital electrical parameters directly from the physical model.

Conclusion and Future Work

While FEMXR provides a powerful tool for resonator design, several areas for improvement remain. Current limitations include the need for more sophisticated strategies for tuning third-party libraries to specific hardware. Furthermore, future iterations of the software aim to incorporate material viscosity (to model energy loss) and thermal effects (to model frequency stability over temperature ranges). Enhancing visualization capabilities for complex mode shapes also remains a priority for improving the tool's utility for design engineers.

Summary updated

Abstract

The finite element analysis of high frequency vibrations of quartz crystal plates is a necessary process required in the design of quartz crystal resonators of precision types for applications in filters and sensors. The anisotropic materials and extremely high frequency in radiofrequency range of resonators determine that vibration frequency spectra are complicated with strong couplings of large number of different vibration modes representing deformations which do not appear in usual structural problems. For instance, the higher-order thickness-shear vibrations usually representing the sharp deformation of thin plates in the thickness direction, expecting the analysis is to be done with refined meshing schemes along the relatively small thickness and consequently the large plane area. To be able to represent the precise vibration mode shapes, a very large number of elements are needed in the finite element analysis with either the three-dimensional theory or the higher-order plate theory, although considerable reduction of numbers of degree-of-freedom (DOF) are expected for the two-dimensional analysis without scarifying the accuracy. In this paper, we reviewed the software architecture for the analysis and demonstrated the evaluation and tuning of parameters for the improvement of the analysis with problems of elements with a large number of DOF in each node, or a problem with unusually large bandwidth of the banded stiffness and mass matrices in comparison with conventional finite element formulation. Such a problem can be used as an example for the optimization and tuning of problems from multi-physics analysis which are increasingly important in applications with excessive large number of DOF and bandwidth in engineering.
BibTeX
@misc{wang2015tuning,
  author = {Ji Wang and Lihong V. Wang and Qiang Sun and Rongxing Wu and Bin Huang and Jianke Du and Wei Xiang},
  title = {{Tuning and Optimizing the Finite Element Analysis with Elements of Large Nodal DOF on a Linux Cluster}},
  year = {2015},
  eprint = {1509.05503},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1509.05503},
}
Figure