VpROM: A novel Variational AutoEncoder-boosted Reduced Order Model for the treatment of parametric dependencies in nonlinear systems

Thomas Simpson, Konstantinos Vlachas, Anthony Garland, Nikolaos Dervilis, Eleni N. Chatzi

TL;DR

This paper introduces VpROM, a novel approach to creating Reduced Order Models (ROMs) for complex, nonlinear systems. By replacing traditional clustering/interpolation methods with a Variational Autoencoder (VAE), VpROM can more accurately capture system behavior across a wide range of parameters and even quantify uncertainty in its predictions.

What Problem Does This Paper Address?

In engineering simulations, especially for complex systems with many parameters or strong nonlinearities, the computational cost of running full-order models (FOMs) can be prohibitive. Reduced Order Models (ROMs) offer a solution by creating faster, lower-dimensional approximations. However, traditional ROM techniques, like Proper Orthogonal Decomposition (POD), struggle when dealing with systems that exhibit drastically different behaviors across different parameter values (e.g., varying material properties or excitation types). To address this, researchers often use libraries of "local" ROMs, each tailored to a specific range of parameters. The challenge then becomes efficiently and accurately mapping a given parameter set to the correct local ROM or its underlying basis. Existing methods for this mapping, such as clustering or interpolation, can be inefficient or inaccurate when the relationship between parameters and local bases is highly nonlinear or complex.

Methods

The core innovation of VpROM lies in its use of a Conditional Variational Autoencoder (cVAE) to learn the mapping between system parameters and the coefficients of local ROM bases.

Here's a breakdown:

  • Projection-Based ROMs: The foundation remains projection-based methods (like POD) to create a low-dimensional subspace. The system's dynamics are then projected onto this subspace.
  • Local Bases: Instead of a single global basis, VpROM utilizes a collection of local bases, each optimized for a specific region of the parameter space.
  • cVAE for Basis Coefficient Generation: The novel part is replacing traditional methods (clustering, interpolation) for selecting or generating these local bases. A cVAE is trained to take system parameters (e.g., material properties, excitation characteristics) as input and generate the coefficients that define the local ROM basis. This allows the VAE to learn complex, nonlinear relationships between parameters and the required basis.
  • Probabilistic Nature: VAEs are generative models that inherently provide a probabilistic output. This allows VpROM to not only predict the local basis coefficients but also quantify the uncertainty associated with these predictions, which can then be propagated to the ROM's response.
  • Hyper-Reduction (Optional): For further computational speedup, VpROM can be combined with hyper-reduction techniques, which further accelerate the evaluation of nonlinear terms in the ROM.

Datasets

The paper validates VpROM on two distinct case studies:

  1. Two-story shear frame with hysteretic links: This is a benchmark simulator featuring a three-dimensional frame with nonlinear nodal couplings exhibiting Bouc-Wen hysteretic behavior. It's used to test the ROM's ability to handle multi-parametric dependencies and nonlinearities. The parameters varied include material properties (Young's modulus) and characteristics of the hysteretic joints, as well as excitation parameters (amplitude, frequency).
  2. NREL 5-MW wind turbine tower: This is a large-scale, complex FE model that includes nonlinear material behavior (isotropic von Mises plasticity) and additional model uncertainties. This case study is used to demonstrate the efficiency and accuracy of VpROM, especially when combined with hyper-reduction, for real-time applications. Parameters varied include excitation amplitude, yield stress, and Young's modulus.

Key Insights and Ideas

  • VAE as a Generative Basis Mapper: The most significant insight is that a VAE can effectively learn the complex, often nonlinear, mapping from system parameters to the coefficients of local ROM bases. This moves beyond simpler interpolation or clustering, which can fail in highly nonlinear regimes.
  • Probabilistic Basis Generation for Uncertainty Quantification: The inherent probabilistic nature of VAEs allows VpROM to provide not just a single prediction for the local basis but a distribution. This is crucial for understanding and quantifying the uncertainty in the ROM's predictions, which is vital for reliable engineering analysis.
  • Improved Accuracy in Complex Regimes: VpROM demonstrates superior accuracy compared to established methods (MACPROM, CpROM) when dealing with systems exhibiting a wide range of dynamic behaviors due to parametric variations and strong nonlinearities.
  • Efficiency Gains with Hyper-Reduction: When combined with hyper-reduction, VpROM achieves significant speedups (up to ~40x compared to FOM), making it suitable for real-time applications.

Key Results

  • Accuracy: VpROM consistently showed lower median and maximum errors in capturing displacement and acceleration time histories compared to MACPROM and CpROM, particularly in challenging parametric regimes. For instance, in the shear frame example, VpROM achieved a median error of 2.20% for displacements, significantly better than MACPROM (2.42%) and CpROM (2.35%).
  • Uncertainty Quantification: The VpROM framework successfully visualized confidence bounds on its predictions, demonstrating its ability to quantify uncertainty.
  • Speedup: The hyper-reduced VpROM (HP-VpROM) achieved substantial speedups, with a speed-up factor of around 35-36x compared to the FOM in the wind turbine case study, while maintaining high accuracy.
  • Robustness: VpROM showed fewer and less severe outliers in its error distributions, indicating a more robust performance across the parameter space.

Limitations and Future Work

  • Training Complexity: The primary limitation acknowledged is the relative complexity of training cVAE models. They involve a large number of hyperparameters (network architecture, layers, activation functions, learning rate) that require careful tuning, often through heuristics like grid search. This is more involved than tuning parameters for clustering or interpolation methods.
  • Future Work: The paper suggests that the probabilistic nature of VpROM can be further leveraged for more comprehensive uncertainty quantification in ROM predictions, going beyond just the basis coefficients. The authors also imply that the framework's robustness and efficiency make it suitable for a wider range of applications in structural dynamics.

Summary updated

Abstract

Reduced Order Models (ROMs) are of considerable importance in many areas of engineering in which computational time presents difficulties. Established approaches employ projection-based reduction such as Proper Orthogonal Decomposition, however, such methods can become inefficient or fail in the case of parameteric or strongly nonlinear models. Such limitations are usually tackled via a library of local reduction bases each of which being valid for a given parameter vector. The success of such methods, however, is strongly reliant upon the method used to relate the parameter vectors to the local bases, this is typically achieved using clustering or interpolation methods. We propose the replacement of these methods with a Variational Autoencoder (VAE) to be used as a generative model which can infer the local basis corresponding to a given parameter vector in a probabilistic manner. The resulting VAE-boosted parametric ROM \emph{VpROM} still retains the physical insights of a projection-based method but also allows for better treatment of problems where model dependencies or excitation traits cause the dynamic behavior to span multiple response regimes. Moreover, the probabilistic treatment of the VAE representation allows for uncertainty quantification on the reduction bases which may then be propagated to the ROM response. The performance of the proposed approach is validated on an open-source simulation benchmark featuring hysteresis and multi-parametric dependencies, and on a large-scale wind turbine tower characterised by nonlinear material behavior and model uncertainty.
BibTeX
@inproceedings{simpson2024vprom,
  author = {Thomas Simpson and Konstantinos Vlachas and Anthony Garland and Nikolaos Dervilis and Eleni N. Chatzi},
  title = {{VpROM: A novel Variational AutoEncoder-boosted Reduced Order Model for the treatment of parametric dependencies in nonlinear systems}},
  booktitle = {Scientific},
  year = {2024},
  doi = {10.1038/s41598-024-56118-x},
}
Figure