A Physics-Guided Machine Learning Model for Predicting Viscoelasticity of Solids at Large Deformation

Area of Science:

• Solid mechanics
• Computational materials science
• Machine learning applications

Background:

• Physics-guided machine learning (PGML) effectively models material constitutive relations by integrating data and physical laws.
• Existing PGML methods excel in time-independent behaviors like elasticity and plasticity but struggle with time-dependent viscoelasticity.
• Accurate modeling of viscoelasticity is challenging, especially with limited experimental data, due to its path and time dependency.

Purpose of the Study:

• To develop a novel physics-guided recurrent neural network (RNN) model for predicting the viscoelastic behavior of solids under large deformations.
• To address the challenge of data scarcity in modeling viscoelasticity by leveraging physics-guided initialization.
• To enable accurate prediction of stress considering time and loading path dependencies.

Main Methods:

• A hybrid GRU-FNN (gated recurrent unit-feedforward neural network) model was developed, accepting time and stretch/strain sequences as input.
• A physics-guided initialization strategy was employed, using numerical data from the generalized Maxwell model for VHB polymers.
• The model was trained using limited experimental data, benefiting from the physics-informed initialization to overcome data scarcity.

Main Results:

• The proposed PGML model successfully predicts the time- and path-dependent viscoelastic behavior of solids at large deformations.
• The physics-guided initialization significantly improved model performance, particularly in scenarios with limited experimental data.
• The GRU-FNN architecture effectively captured the complex stress-strain-time relationships inherent in viscoelastic materials.

Conclusions:

• Physics-guided recurrent neural networks offer a powerful framework for modeling complex viscoelastic phenomena, even with scarce experimental data.
• The integration of physical knowledge through initialization enhances the robustness and accuracy of machine learning models for material science.
• This approach advances the capability to predict constitutive relations for viscoelastic materials, with implications for material design and performance simulation.