A generic physics-informed neural network-based constitutive model for soft biological tissues.

Area of Science:

• Computational mechanics
• Biomaterials science
• Machine learning applications

Background:

• Constitutive modeling is essential for analyzing the mechanical behavior of biological soft tissues.
• Machine learning (ML) techniques show promise in establishing direct relationships between strain and stress for constitutive modeling.

Purpose of the Study:

• To develop a novel, generic physics-informed neural network material (NNMat) model for constitutive modeling of biological soft tissues.
• To implement a hierarchical learning strategy for characterizing general material properties and individual subject responses.

Main Methods:

• Proposed a novel neural network structure with class and subject parameter sets for hierarchical learning.
• Incorporated skip connections and a convexity constraint to ensure physical relevance.
• Trained, cross-validated, and tested the NNMat model using biaxial testing data from 63 ascending thoracic aortic aneurysm tissue samples.

Main Results:

• The NNMat model demonstrated significantly better fitting performance (R² = 0.9632) compared to the Holzapfel-Gasser-Ogden model (R² = 0.9019).
• The NNMat model also showed superior testing performance (R² = 0.9471) versus the Holzapfel-Gasser-Ogden model (R² = 0.8556).
• The model allows direct adoption for new subjects by only re-training subject-specific parameters.

Conclusions:

• The developed NNMat model provides a convenient and generalizable methodology for constitutive modeling of biological soft tissues.
• The physics-informed neural network approach offers improved accuracy and efficiency over traditional expert-constructed models.