Eig-PIELM: A Mesh-Free Approach for Efficient Eigen-Analysis with Physics-Informed Extreme Learning Machines

1. Problem Statement

The paper addresses the computational challenge of solving linear eigenvalue problems (e.g., $Lv = \lambda v$) in continuum mechanics and physics. These problems are fundamental to analyzing natural frequencies and mode shapes in structural vibrations, acoustics, and electromagnetics.

Key Challenges in Existing Methods:

• Classical Methods (FEM/FDM): Require mesh generation, which is computationally expensive for complex geometries and parametric studies.
• Physics-Informed Neural Networks (PINNs): While mesh-free, they suffer from non-convex optimization landscapes, sensitivity to hyperparameters, high training costs due to backpropagation, and spectral bias (difficulty capturing high-frequency modes).
• Standard Physics-Informed Extreme Learning Machines (PIELM): While PIELM offers a fast, single-step linear solution by freezing hidden layer weights, it fails for eigenvalue problems. In eigenproblems, the residual depends on both the unknown eigenvalue ($\lambda$) and the coefficients, breaking the linearity required for a single-step solve. Existing adaptations often require iterative loops or assume $\lambda$ is known, negating PIELM's efficiency.

2. Methodology: The Eig-PIELM Framework

The authors propose Eig-PIELM, a novel framework that reformulates the eigenvalue problem into a single symmetric generalized eigenvalue problem, preserving the "one-shot" computational advantage of PIELM.

Core Mechanisms:

1. Ansatz and Discretization:

   • The solution $u(x)$ is approximated using a single-hidden-layer neural network with frozen random (or polynomial) features: $\hat{u}(x) = \sum \beta_k \sigma(w_k x + b_k)$.
   • The problem is discretized using collocation points in the interior domain ($\Omega$) and on the boundary ($\partial\Omega$).

2. Algebraic Projection for Exact Boundary Enforcement:

   • Unlike PINNs which use penalty terms to approximate boundary conditions, Eig-PIELM enforces boundary conditions exactly.
   • It introduces a linear transformation matrix $T_b$ that maps the general coefficient space to a boundary-admissible subspace.
   • By reparameterizing coefficients as $\beta = T_b y$, the boundary residuals become identically zero ($B\hat{u} = 0$) without needing penalty weights or Lagrange multipliers.

3. Derivation of the Generalized Eigenproblem:

   • The interior residual is defined as $r = L\hat{u} - \lambda \hat{u}$.
   • Minimizing the least-squares loss $J(\beta, \lambda)$ with respect to both the coefficients and $\lambda$ yields a coupled system.
   • Through the algebraic projection, the skew-symmetric parts of the system matrices vanish, collapsing the coupled equations into a single symmetric generalized eigenvalue problem:
     $$S_{red} y = \lambda G_{red} y$$
   • Here, $S_{red}$ and $G_{red}$ are reduced matrices derived from the differential operator and basis functions, respectively. $G_{red}$ is symmetric positive definite.

4. Solution:

   • The system is solved using standard, non-iterative linear algebra routines (e.g., QR algorithm or Cholesky decomposition) to obtain all eigenvalues ($\lambda$) and eigenvectors ($y$) simultaneously.
   • The mode shapes are reconstructed via $\hat{u}(x) = \phi(x)^T T_b y$.

3. Key Contributions

• Novel Framework: First extension of PIELM to eigenvalue problems that retains the single-step, non-iterative nature of the method.
• Exact Boundary Conditions: Utilization of algebraic projection to satisfy boundary conditions exactly, eliminating the need for penalty parameters and the associated tuning difficulties.
• Simultaneous Spectrum Computation: The method computes the entire spectrum of eigenvalues and mode shapes in a single linear solve, avoiding the need for separate optimization runs for each mode.
• Mesh-Free Efficiency: The approach is entirely mesh-free, relying on collocation points, and avoids backpropagation, making it significantly faster than PINNs.

4. Numerical Results and Validation

The framework was tested on two primary benchmark problems using MATLAB on an Intel Core i9 processor.

A. Transverse Vibration of Euler-Bernoulli Beams

• Cases: Simply supported, Fixed-Fixed, and Cantilever (Fixed-Free) beams.
• Basis: Bernstein polynomials.
• Performance:
  • Accuracy: Achieved near machine precision for fundamental modes (relative errors $\sim 10^{-16}$ to $10^{-20}$) and high accuracy for higher-order modes (errors $\sim 10^{-9}$ to $10^{-12}$).
  • Speed: Computed the first five modes for various configurations in 0.16 to 0.199 seconds.
  • Observation: The method successfully reproduced complex, oscillatory higher-order mode shapes without spectral bias.

B. 2D Acoustic Cavity (Helmholtz Equation)

• Case: Rectangular domain with rigid walls (Neumann boundary conditions).
• Performance:
  • Accuracy: Relative errors in the range of $10^{-8}$ to $10^{-7}$ for the first five modes.
  • Speed: Full eigenspectrum computed in 0.244 seconds.
  • Observation: Demonstrated robustness in 2D elliptic eigenproblems with doubly indexed analytical solutions.

5. Significance and Conclusion

The Eig-PIELM framework represents a significant advancement in mesh-free computational mechanics. By resolving the non-linearity inherent in eigenproblems within the PIELM architecture, it offers a robust, accurate, and extremely fast alternative to traditional FEM and iterative PINNs.

Key Implications:

• Parametric Studies: The speed and non-iterative nature make it ideal for design optimization and parametric sensitivity analysis where thousands of eigenvalue evaluations are required.
• High-Frequency Analysis: The method overcomes the spectral bias common in neural network methods, accurately capturing high-frequency modes.
• Simplicity: The elimination of penalty weights and backpropagation simplifies implementation and reduces hyperparameter tuning.

The authors conclude that Eig-PIELM fills a critical gap in eigenanalysis, particularly for mechanical, acoustic, and electromechanical systems requiring rapid frequency spectrum analysis. Future work aims to extend this to nonlinear eigenvalue problems and large-scale industrial applications.