Experimental Learning of a Hyperelastic Behavior with a Physics-Augmented Neural Network

TL;DR

This paper demonstrates that Physics-Augmented Neural Networks trained on real experimental data can effectively model hyperelastic material behavior, outperforming traditional models like Neo-Hookean in accuracy and generalization.

Contribution

The study introduces a method for training PANN with actual experimental data to model hyperelasticity, showing improved performance over conventional models.

Findings

PANN accurately captures hyperelastic behavior beyond training loads.

Training on experimental data outperforms the Neo-Hookean model.

Model generalizes well to independent experimental data.

Abstract

The recent development of Physics-Augmented Neural Networks (PANN) opens new opportunities for modeling material behaviors. These approaches have demonstrated their efficiency when trained on synthetic cases. This study aims to demonstrate the effectiveness of training PANN using real experimental data for modeling hyperelastic behavior. The approach involved two uni-axial experiments equipped with digital image correlation and force sensors. The tests achieved axial deformations exceeding 200% and presented non-linear responses. Twenty loading steps extracted from one experiment were used to train the PANN. The model architecture was optimized based on results from a validation dataset, utilizing equilibrium gap loss computed on six loading steps. Finally, 544 loading steps from the first experiment and 80 steps from a second independent experiment were used for testing purposes. The PANN model effectively captured the hyperelastic behavior across and beyond the training loads, showing superior performance compared to the standard Neo-Hookean model when assessed using various evaluation metrics. Training PANN with experimental mechanical data shows promising results, outperforming traditional modeling approaches.