Spectrum and Physics-Informed Neural Networks (SaPINNs) for Input–State–Parameter Estimation in Dynamic Systems Subjected to Natural Hazards-Induced Excitation

Antonina M. Kosikova, Apostolos F. Psaros, Andrew W. SmythSSRN 2025 · DOI

TL;DR

This paper introduces Spectrum and Physics-Informed Neural Networks (SaPINNs), a novel framework that enhances traditional Physics-Informed Neural Networks (PINNs) by incorporating spectral characteristics of natural hazard excitations. SaPINNs demonstrate superior performance in estimating unknown input forces, system states, and parameters for dynamic systems under complex excitations like earthquakes and windstorms, offering more accurate and physically consistent results with quantified uncertainty.

What Problem Does This Paper Address?

System identification, particularly in structural health monitoring (SHM), often faces the challenge of estimating system states and parameters when the external excitation forces are unknown or only partially characterized. This problem is exacerbated in dynamic systems subjected to natural hazards like earthquakes and windstorms, where excitations are highly transient, nonlinear, and spatially distributed. Traditional methods like finite element methods (FEM) and Bayesian techniques can be computationally intensive and data-hungry. While standard PINNs can embed physics into neural networks, they struggle with unknown, non-stationary excitations and can converge to non-unique solutions. This paper addresses the need for a more robust and efficient method to accurately identify system dynamics under such complex and uncertain loading conditions.

Methods

The core innovation is the development of Spectrum and Physics-Informed Neural Networks (SaPINNs). This framework builds upon standard PINNs by integrating spectral information of natural hazards into the loss function. Specifically, the Spectral Representation Method (SRM) is used to generate plausible realizations of stationary stochastic processes based on a target power spectral density (PSD). The unknown excitation force is then modeled as a realization of this process, parameterized by phase angles. These phase angles, along with system parameters and neural network weights, are jointly optimized. This approach introduces a spectral constraint, acting as an inductive bias that guides the network towards physically consistent solutions. To quantify uncertainty, the SaPINNs are embedded within a Deep Ensemble (DEns) architecture, training multiple networks independently to approximate the predictive distribution.

Datasets

The paper uses synthetic datasets generated to simulate the dynamic responses of linear and nonlinear systems under various excitation scenarios. These include:

  • Sinusoidal excitation: For a single-degree-of-freedom (SDoF) system.
  • Thunderstorm wind loading: Modeled using an empirically derived Empirical Power Spectral Density (EPSD) model.
  • Seismic excitation:
    • Synthetic earthquake ground motion simulated from a Kanai-Tajimi EPSD.
  • The North-South component of the 1940 El Centro earthquake.

Key Insights and Ideas

  • Spectral information as a powerful inductive bias: Incorporating the spectral characteristics of natural hazards into the PINN loss function significantly constrains the solution space, guiding the network towards more physically plausible and accurate estimates, especially for unknown excitation forces.
  • Addressing ill-posedness: The problem of input-state-parameter estimation with unknown excitations is inherently ill-posed. SaPINNs, by leveraging spectral priors, effectively reduce the dimensionality of the search space and mitigate issues of non-uniqueness and identifiability.
  • Deep Ensembles for robust uncertainty quantification: Using DEns allows for the estimation of epistemic uncertainty, providing a measure of confidence in the predictions, which is crucial for real-world applications where data and models are imperfect.
  • Limitations of standard PINNs: The study highlights that a low training loss in standard PINNs does not necessarily guarantee accurate or physically consistent parameter estimates, especially in complex scenarios with unknown excitations.
  • Observability is key: The paper emphasizes the importance of observability analysis to ensure that the system and its parameters can be uniquely identified from the available measurements, especially when dealing with unknown inputs.

Key Results

  • Superior input force reconstruction: SaPINNs consistently outperformed standard PINNs in reconstructing unknown excitation forces, particularly for complex wind and seismic events.
  • Accurate parameter estimation: SaPINNs achieved more accurate estimates of system parameters (e.g., stiffness scaling factors) with lower uncertainty compared to PINNs.
  • Improved state estimation: While both models could recover displacement responses, SaPINNs provided more physically consistent trajectories, especially in nonlinear systems.
  • Quantified uncertainty: The DEns approach provided reliable estimates of uncertainty for the inferred quantities, demonstrating that SaPINNs lead to reduced uncertainty compared to standard PINNs.
  • Faster convergence: In some scenarios, SaPINNs showed faster convergence to a stable loss compared to PINNs.

Limitations and Future Work

  • Dependence on measurement quality and spectral content: The performance of SaPINNs is influenced by the quality of available measurements and the accuracy of the assumed spectral content of the excitation.
  • Model capacity and training time: Scenarios with high-frequency content or significant model misspecification may require deeper architectures and wider layers, increasing training time and the risk of overfitting.
  • Extension to complex MDoF systems: The framework could be extended to more complex multi-degree-of-freedom (MDoF) systems, but this might introduce greater discrepancies between models and real systems.
  • Incorporating modeling error: For more complex systems, explicitly accounting for modeling error using auxiliary processes like Gaussian Processes is recommended.
  • Broader applications: The framework holds promise for applications beyond civil structures, such as aerospace engineering, for structural health monitoring and fatigue-life assessment.

Summary updated

BibTeX
@misc{kosikova2025spectrum,
  author = {Antonina M. Kosikova and Apostolos F. Psaros and Andrew W. Smyth},
  title = {{Spectrum and Physics-Informed Neural Networks (SaPINNs) for Input–State–Parameter Estimation in Dynamic Systems Subjected to Natural Hazards-Induced Excitation}},
  year = {2025},
  doi = {10.2139/ssrn.5606076},
  howpublished = {SSRN},
}
Figure