Solving digital image correlation with neural networks constrained by strain-displacement relations.

Area of Science:

• Computational mechanics
• Digital image correlation (DIC)
• Machine learning applications

Background:

• Existing supervised neural network methods for DIC treat the process as a black box, ignoring physical constraints between displacement and strain fields.
• This black-box approach leads to suboptimal accuracy, sometimes performing worse than traditional Subset-DIC methods.
• A need exists for more physically informed deep learning models in DIC analysis.

Purpose of the Study:

• To develop a deep learning model for digital image correlation (DIC) that incorporates physical strain-displacement relationships.
• To improve the accuracy of deformation field calculations by considering physical constraints during neural network training.
• To evaluate the performance of the proposed physically constrained model against traditional methods.

Main Methods:

• A novel deep learning architecture was designed, integrating known strain-displacement relations directly into the neural network.
• The model training considers errors in both displacement and strain fields, guided by the physical constraints.
• The back-propagation process was derived, and the solution was implemented using Python for simulation and experimental validation.

Main Results:

• The proposed deep learning model, incorporating physical constraints, demonstrated significantly improved prediction accuracy compared to existing black-box methods.
• Evaluation through simulations and real digital image correlation experiments confirmed the enhanced performance.
• The inclusion of strain-displacement relations proved crucial for boosting the accuracy of deformation field predictions.

Conclusions:

• Integrating physical constraints, specifically strain-displacement relations, into neural networks is an effective strategy for enhancing DIC accuracy.
• The developed deep learning model offers a more robust and accurate approach to analyzing deformation fields from image data.
• This physically informed deep learning method represents a significant advancement over conventional black-box neural network applications in DIC.