Can machine learning accelerate soft material parameter identification from complex mechanical test data?

Area of Science:

• Biomechanics
• Materials Science
• Computational Science

Background:

• Identifying soft material parameters typically requires complex mechanical tests and computationally expensive inverse strategies.
• Previous methods involving iterative finite element solutions are often prohibitively time-consuming.

Purpose of the Study:

• To investigate machine learning (ML) approaches for accelerating the identification of constitutive parameters in soft materials.
• To evaluate two ML strategies: replacing finite element (FE) simulations with ML metamodels and using a standalone neural network (NN) to directly predict parameters.

Main Methods:

• Developed ML-based metamodels (Gaussian Process Regression, Neural Networks) to replace FE simulations within optimization frameworks.
• Trained and evaluated a standalone Neural Network to directly predict material parameters from experimental data.
• Tested approaches on simple shear experiments of blood clot and simple shear/uniaxial loading of right ventricular myocardium.

Main Results:

• Replacing FE simulations with ML metamodels significantly accelerated parameter identification for both blood clot and myocardium.
• Metamodels yielded excellent results for the homogeneous blood clot and satisfying results for anisotropic myocardium.
• Direct prediction using a standalone NN provided unsatisfying results, particularly for the heterogeneous myocardium.

Conclusions:

• ML-based metamodels offer a viable and efficient alternative to traditional FE simulations for soft material parameter identification.
• The study provides a benchmark dataset for applying ML to soft tissue biomechanics, demonstrating ML's potential to accelerate analysis from complex mechanical data.