Neural-VSI: Variational System Identification of Structural Parameter Fields in High-Order PDEs

Xuyang Li, Mahdi Masmoudi, Rami Gharbi, Nizar Lajnef, Vishnu Boddeti, John Harlim, Romit Maulik

Keywords: Physics-Informed Machine Learning, Parameter Estimation, Weak Form, System Identification, Structural Health Monitoring

TL;DR: Neural-VSI utilizes local weak forms with pre-computed matrix operators to identify spatially varying parameters in high-order PDEs, replacing automatic differentiation to achieve a 200 times speedup compared to the strong form..

Abstract:

Structural parameter identification in high-order partial differential equations (PDEs), such as the Euler-Bernoulli beam equation, remains challenging due to the computation of high-order derivative operators, particularly when structural parameters vary spatially.
Current physics-informed machine learning approaches, including Physics-Informed Neural Networks (PINNs) and Universal Differential Equations (UDEs), typically require expensive automatic differentiation (AD) or adjoint calculations.
These methods often fail when measurement data is noisy or boundary conditions (BCs) are unknown.
This paper proposes Neural-VSI, a variational framework that parameterizes unknown fields with neural networks and reformulates the inverse problem using local variational forms with pre-computed integration weight matrices.
Our method bypasses the heavy computational overhead of AD, allowing for the estimation of hard-to-measure structural parameter fields without utilizing boundary condition information.
Experiments on the vibration of beams with two unknown distributed parameters demonstrate that this approach achieves a speedup of over 200 times compared to strong-form baselines while maintaining robust identification accuracy.

Journal Opt In: No, I do not wish to participate

Journal Corresponding Email: [e-mail]

Submission Number: 123