Machine Learning Identification of Piezoelectric Properties.
Mariana Del Castillo1, Nicolás Pérez1
1Faculty of Engineering, University of the Republic (UdelaR), 11300 Montevideo, Uruguay.
Materials (Basel, Switzerland)
|June 2, 2021
Summary
Machine learning accurately identifies piezoelectric model parameters, overcoming simulation limitations. This fast approach aids manufacturers and end-users working with specific materials and geometries.
Area of Science:
- Materials Science
- Computational Mechanics
- Artificial Intelligence
Background:
- Numerical simulations accurately model piezoelectric elements but require precise material parameters.
- Identifying these parameters is a persistent challenge for manufacturers and end-users.
- Existing parameter identification techniques have limitations.
Purpose of the Study:
- To present a machine learning (ML) approach for determining key piezoelectric model parameters.
- To validate the ML model's accuracy using finite element simulations.
- To assess the approach's speed and suitability for industrial applications.
Main Methods:
- A neural network was trained using approximately one million finite element simulations.
- Key sensitive parameters (c11, c13, c33, c44, e33) were predicted.
- A PZT 27 piezoelectric ceramic (20 mm diameter, 2 mm thickness) dataset was used for initial training.
Main Results:
- The ML approach accurately predicted piezoelectric parameters, with errors under 0.6% in the worst case.
- The trained neural network provides extremely fast parameter identification post-training.
- The method demonstrated high fidelity in reproducing original parameter values.
Conclusions:
- Machine learning offers a highly accurate and efficient solution for piezoelectric model parameter identification.
- This approach significantly reduces the limitations posed by parameter uncertainty in simulations.
- The method is particularly beneficial for manufacturers and end-users with consistent material and geometric requirements.
