A plausible Parametrization of Modal Basis for Dynamical Systems Analysis

Technical Summary: A Plausible Parametrization of Modal Basis for Dynamical Systems Analysis

Problem Statement
In solid dynamics, the modal basis—comprising eigenvalues (natural frequencies) and eigenvectors (mode shapes)—is fundamental for vibration analysis, structural optimization, and control-oriented design. Knowledge of this basis allows engineers to predict resonance, avoid critical frequencies, and tailor dynamic responses. However, for realistic engineering systems, determining the modal basis requires solving large-scale eigenvalue problems, typically via the Finite Element Method (FEM). While accurate, these computations are computationally expensive, particularly in parametric design and optimization contexts where geometry, material properties, or boundary conditions vary, necessitating repeated eigenvalue solves.

Existing Model Order Reduction (MOR) techniques, such as Proper Orthogonal Decomposition (POD) and Reduced Basis (RB) methods, attempt to alleviate this burden by exploiting low-dimensional solution structures. However, these methods often rely on linear reduced spaces, which may struggle to capture strongly nonlinear modal dependencies induced by complex parametric variations. Furthermore, standard Deep Learning approaches, while powerful for surrogate modeling, often suffer from overfitting, lack of interpretability, and the generation of non-physical "hallucinated" features when applied to physics-based problems.

Methodology
The authors propose a novel framework utilizing a Rank Reduction AutoEncoder (RRAE) to parametrize the modal basis. The methodology addresses two primary challenges: data preprocessing (mode alignment) and the construction of a nonlinear parametric surrogate.

1. Data Preprocessing (Mode Matching and Alignment):
   Eigenvectors are defined only up to a scaling factor (sign ambiguity) and their ordering can shift when system parameters change. To ensure consistent learning, the authors employ a pre-processing step using an $L_1$-regularized minimization approach (Lasso solver). This technique matches disturbed modes to nominal modes by finding a sparse weight vector that minimizes the $L_2$ norm between the nominal mode and a linear combination of disturbed modes. The sign of the resulting weight indicates whether a mode needs to be flipped, and the magnitude identifies the correct permutation, effectively re-ordering and aligning the dataset.

2. Rank Reduction AutoEncoder (RRAE) Architecture:
   The core of the proposed method is an autoencoder where the latent space is constrained via a truncated Singular Value Decomposition (SVD). Instead of allowing the latent vector to take an arbitrary form, the RRAE restructures it into a matrix whose SVD is truncated to a predefined rank $k_{max} \ll L$. This forces the latent space to capture only the dominant features of the data, mitigating overfitting and spurious predictions.

3. Parametric Surrogate Construction:
   The framework extends the RRAE to handle parametric inputs ($p$) through a structured learning strategy:

   • Parametric Network: A neural network takes design parameters $p$ as input and predicts the low-rank latent representation ($V\alpha$) corresponding to the first eigenvector ($\phi_1$).
   • Mode Coupling Networks: $N-1$ supplementary neural networks are introduced. Each network $i$ receives the latent representation predicted by the parametric network and outputs the $i$-th eigenvector ($\phi_i$) for $i > 1$.
   • Joint Training: All networks are trained jointly to minimize a composite loss function consisting of:
     • Reconstruction Loss ($L_{rec}$): Ensures the RRAE accurately reconstructs the first eigenvector from the latent space.
     • Parametric Loss ($L_{param}$): Links the design parameters to the regularized low-rank latent space.
     • Eigenvector Prediction Loss ($L_{eigv}$): Guides the latent representation to accurately predict all remaining eigenvectors via the supplementary networks.

Key Contributions

• RRAE for Modal Basis: The application of a Rank Reduction AutoEncoder, constrained by truncated SVD, specifically for the parametrization of eigenvectors in solid dynamics.
• Robust Preprocessing: A systematic $L_1$-based strategy to resolve sign and ordering ambiguities in eigenvector datasets, a critical step often overlooked in data-driven modal analysis.
• Nonlinear Parametric Framework: A coupled architecture that learns a shared low-dimensional latent space for the first mode and propagates this information through dedicated networks to reconstruct the entire modal basis, capturing nonlinear dependencies that linear MOR methods may miss.
• Validation: The methodology is validated on both 1D (bar) and 2D (bending plate) problems with varying material fractions and stiffness properties.

Results
The proposed architecture was tested on:

• 1D Bar Problem: A bar with two material domains defined by a fraction parameter $g$. Using 2,000 Latin Hypercube samples, the model successfully predicted the first five eigenvectors for unseen test data.
• 2D Plate Problem: A cantilever plate divided into two domains with varying Young's moduli ($E_1, E_2$) and a geometric fraction $g$. The model utilized 500 data points and 2D Convolutional Neural Networks.

The results demonstrated that the parametric surrogate correctly predicted the group of eigenvectors. In the 2D plate case, while the total test error was reported as 16%, the authors clarify that this aggregates errors across all modes; the individual error for each mode remained below 4%. The training and test loss curves indicated convergence, confirming the model's ability to generalize across the parametric space.

Significance and Claims
The paper claims that the RRAE framework provides an efficient surrogate model for parametric modal analysis, offering a viable alternative to repeated, expensive FEM eigenvalue solves. By constraining the latent space through truncated SVD, the method effectively captures dominant physical features while avoiding the non-physical representations common in unconstrained deep learning models. The authors position this work as a step toward efficient system optimization design in solid dynamics. They note that while the current work focuses on 1D and 2D problems for simplification, the methodology is directly applicable to 3D solid mechanics problems, which they identify as a perspective for future work. The paper maintains a modest tone, focusing on the validation of the architecture's ability to parametrize eigenvectors rather than claiming universal superiority over all existing MOR techniques.