NeuralSI: Structural Parameter Identification in Nonlinear Dynamical Systems

Xuyang Li, Hamed Bolandi, Talal Salem, Nizar Lajnef, Vishnu Naresh Boddeti

Introduction

Structural Health Monitoring (SHM) is a critical field focused on assessing the integrity and remaining useful life of infrastructure like bridges, buildings, and airframes. A persistent challenge in SHM is the discrepancy between a structure's original design and its actual physical state after years of environmental wear or damage. Identifying "structural parameters"—such as the spatial distribution of material stiffness (modulus) and energy loss (damping)—is essential for accurate performance prediction. However, traditional methods struggle with the high-order nonlinear partial differential equations (PDEs) governing these systems, especially when sensor data is sparse.

This paper introduces NeuralSI, a hybrid framework that bridges the gap between data-driven neural networks and classical physics. By embedding neural networks within the governing physical equations, NeuralSI identifies unknown structural parameters with unprecedented accuracy and generalizability.

The NeuralSI Framework

The core innovation of NeuralSI is its "physics-in-the-loop" architecture. Unlike standard deep learning models that attempt to map inputs directly to outputs (e.g., predicting displacement from time), NeuralSI is tasked with learning the underlying physical properties of the structure.

Architecture and Physics Integration

  1. Parameter Prediction: A Feed-Forward Neural Network takes spatial coordinates as input and outputs the estimated modulus coefficient ($P$) and damping coefficient ($C$). To improve the network's understanding of spatial relationships, the model incorporates positional encodings similar to those used in Transformer architectures.
  2. Differentiable Physics Solver: The predicted parameters are plugged into a discretized Euler-Bernoulli beam equation. Using the Method of Lines, the PDE is converted into a system of Ordinary Differential Equations (ODEs).
  3. End-to-End Optimization: Leveraging the Julia programming language's differentiable programming ecosystem, the framework backpropagates the error from the final displacement prediction through the physics solver to update the neural network weights. This ensures the identified parameters are mathematically consistent with the laws of structural dynamics.

Experimental Setup

The authors evaluated NeuralSI using a simulated nonlinear beam vibration dataset. The setup involved a simply supported aluminum beam (40cm x 5cm x 0.5cm) with synthetic ground truth parameters: a sinusoidally varying modulus and a ramp-function damping profile across its length.

The model was trained on a sparse displacement distribution ($u(x, t)$) consisting of 16 sensor locations over 160 time steps. NeuralSI was benchmarked against two primary baselines:

  • Deep Neural Networks (DNN): A purely data-driven approach.
  • Physics-Informed Neural Networks (PINN): A common baseline that regularizes neural networks with physical constraints.

Key Insights and Results

Superior Accuracy and Extrapolation

NeuralSI achieved a breakthrough in precision, reducing interpolation and extrapolation errors by two to five orders of magnitude compared to DNN and PINN baselines. While standard models often "hallucinate" unphysical vibrations when predicting outside their training range, NeuralSI remained highly stable. In temporal extrapolation tests—predicting the beam's movement twice as far into the future as the training data allowed—NeuralSI maintained a peak error of only ~0.3%, while the baselines diverged significantly.

Data Efficiency and Speed

The framework proved remarkably robust to data scarcity, performing well even when trained on only 20% of the available sensor data. This is particularly relevant for real-world applications where sensor coverage is often limited by cost or accessibility. Additionally, by utilizing optimized solvers (ModelingToolkit.jl), the authors achieved a 40x speedup in inference compared to their initial training configurations.

The Challenge of Damping

The study highlighted that identifying the damping coefficient ($C$) is significantly more difficult than identifying the modulus ($P$). Because the system's overall response is less sensitive to minor fluctuations in damping, the "signal" is harder for the network to isolate, marking an area for further refinement in structural modeling.

Conclusion and Future Work

NeuralSI demonstrates that learning the cause (physical parameters) rather than just the effect (displacement) leads to models that are both more accurate and more physically grounded. By using the PDE as a massive regularizer, NeuralSI avoids the common pitfalls of "black box" models in engineering.

Future research will focus on extending the framework to handle spatiotemporal variations (where parameters change over both space and time) and scaling the computation to handle complex 3D geometries. Validating the framework against noisy, real-world experimental data remains the next critical step toward deploying NeuralSI in active structural health monitoring systems.

Summary updated

BibTeX
@inproceedings{li2022neuralsi,
  author = {Xuyang Li and Hamed Bolandi and Talal Salem and Nizar Lajnef and Vishnu Naresh Boddeti},
  title = {{NeuralSI: Structural Parameter Identification in Nonlinear Dynamical Systems}},
  booktitle = {Workshops @ ECCV},
  year = {2022},
  url = {https://openreview.net/forum?id=uCbH0WcFau},
}
Figure