A Machine Learning-Based Surrogate Finite Element Model for Estimating Dynamic Response of Mechanical Systems

Ali Hashemi, Jinwoo Jang, Javad BeheshtiIEEE Access 2023 · DOI

Introduction

Structural Health Monitoring (SHM) and safety evaluations rely heavily on predicting how mechanical structures respond to external forces. Finite Element Analysis (FEA) is the gold standard for these predictions; however, it is computationally intensive, often requiring hours or days for complex, time-varying simulations. This computational "expense" creates a significant bottleneck for real-time applications and iterative design processes. This paper proposes a solution: a machine learning-based surrogate model that approximates heavy FEA solvers with high precision and minimal processing time.

Methodology

The researchers developed a pipeline that bridges physics-based modeling and supervised learning. The core of the study involves a 31-bar planar truss structure, modeled in ANSYS to provide high-fidelity ground truth data.

The 31-bar planar truss structure used to generate the FEA ground truth data for training.

Four machine learning architectures were evaluated to determine the most effective surrogate:

  1. Deep Neural Networks (DNNs): A 9-hidden-layer network using ReLU activation and the Adam optimizer.
  2. Ensemble Tree Methods: Random Forest (RF), AdaBoost, and Extreme Gradient Boosting (XGBoost).

The models were trained to predict dynamic strain responses at 14 nodal locations based on external excitation loads and specific time steps. Hyperparameters for each model were optimized using GridSearch to ensure peak performance.

Dataset and Training

The study utilized a synthetic dataset of over 1.5 million data points generated via FEA. The dataset included 16 attributes: time instances, applied external loads, and 14 nodal strain values. The data was split into an 80% training set (1.22 million points) and a 20% testing set (307,000 points). Z-score normalization was applied to ensure the model could handle the varying magnitudes of input attributes across different directions, specifically handling the X and Y coordinate variances.

Results and Discussion

The results revealed a clear advantage for ensemble tree methods over deep learning for this specific task. XGBoost emerged as the superior model, achieving the highest precision across all test scenarios.

Performance comparison showing the high correlation between predicted and actual values for the surrogate models.

Key performance metrics included:

  • High Accuracy: All three decision tree models achieved $R^2$ values exceeding 0.98, meaning they captured 98% of the data variance.
  • XGBoost Dominance: XGBoost recorded the lowest Mean Absolute Error (MAE) of 8.55 in the X-direction and 9.89 in the Y-direction.
  • DNN Limitations: The DNN model, while functional, lagged behind with an $R^2$ of approximately 0.92 and a significantly higher MAE (14.31 in the X-direction).

The surrogate models demonstrate a "black-box" advantage, estimating response variables without solving complex differential equations of motion at every step. This allows for near real-time estimation of structural states, which is impossible with traditional iterative FEA solvers.

Conclusion

This research successfully demonstrates that machine learning surrogates, particularly XGBoost, can effectively replace traditional FEA for dynamic response estimation in 2D truss structures. While the current study is limited to 2D systems, the authors acknowledge that scaling to 3D structures will require significantly larger datasets. Future work will focus on scaling these models and investigating the optimal placement of "virtual sensors" to improve the perception of dynamic mechanical systems in real-time industrial operations.

Summary updated

BibTeX
@article{hashemi2023machine,
  author = {Ali Hashemi and Jinwoo Jang and Javad Beheshti},
  title = {{A Machine Learning-Based Surrogate Finite Element Model for Estimating Dynamic Response of Mechanical Systems}},
  journal = {IEEE Access},
  year = {2023},
  doi = {10.1109/access.2023.3282453},
}
Figure