Physics-Informed Neural Network Based Digital Image Correlation Method

Background

Deep Learning-based Digital Image Correlation (DL-DIC) approaches take advantages such as pixel-wise calculation in a full-automatic manner without user's input and improved accuracy in non-uniform deformation measurements. However, DL-DIC still faces accuracy limitations due to the lack of high-precision real-world training data in supervised-learning methods and the need for smoothing noisy solutions in unsupervised-learning methods.

Objective

This paper proposes a DIC solution method based on Physics-Informed Neural Networks (PINN), called PINN-DIC, to address deformation measurement challenges of current DL-DIC in practical applications.

Methods

PINN-DIC utilizes a fully connected neural network, with regularized spatial coordinate field as input and displacement field as output. It applies the photometric consistency assumption as a physical constraint, using grayscale differences between predicted and actual deformed images to construct a loss function for iterative optimization of the displacement field. Additionally, a warm-up stage is designed to assist in iterative optimization, allowing PINN-DIC to achieve high accuracy in analyzing both uniform and non-uniform displacement fields.

Results

PINN-DIC, validated through simulations and real experiments, not only maintained the advantages of other DL-DIC methods but also demonstrated superior performance in achieving higher accuracy than conventional unsupervised DIC and handling irregular boundaries with adjusting the input coordinate field.

Conclusions

PINN-DIC is an unsupervised method that takes a regularized coordinate field (instead of speckle images) as input and achieves higher accuracy in deformation field results with a simple network. It introduces a novel approach to DL-DIC, enhancing performance in complex measurement scenarios.