Automated model discovery for human brain using Constitutive Artificial Neural Networks.

Area of Science:

• Computational mechanics
• Biophysics
• Machine learning

Background:

• Understanding the mechanical behavior of the brain is crucial due to its vulnerability.
• Existing constitutive models for soft tissues are numerous, and selection is often subjective and experience-dependent.
• There is a need for an objective and automated method for selecting appropriate constitutive models.

Purpose of the Study:

• To develop a novel strategy for the simultaneous discovery of constitutive models and their parameters for soft biological tissues.
• To create a Constitutive Artificial Neural Network (CANN) capable of automated model discovery.
• To challenge the conventional approach of pre-selecting a model and then fitting parameters.

Main Methods:

• Integrated thermodynamics and machine learning to build a Constitutive Artificial Neural Network (CANN).
• Designed the CANN by reverse-engineering functional building blocks generalizing existing constitutive models.
• Constrained the network architecture and functions to ensure thermodynamic consistency, objectivity, symmetry, and polyconvexity.

Main Results:

• The CANN autonomously discovered the optimal model and parameters for human gray and white matter from over 4000 possibilities.
• The discovered models accurately characterized brain tissue behavior under tension, compression, and shear.
• Network weights translated into physically meaningful parameters, yielding specific shear moduli for different brain regions.

Conclusions:

• Constitutive Artificial Neural Networks (CANNs) offer a paradigm shift in soft tissue modeling, moving from manual model selection to automated discovery.
• This approach provides an objective and efficient method for characterizing complex material behaviors.
• The developed CANN framework has broad potential applications in biomechanics and materials science.