Discovering neural elastoplasticity from kinematic observations.

Area of Science:

• Computational Materials Science
• Solid Mechanics
• Machine Learning

Background:

• Accurate material models are crucial for high-fidelity predictions but often require extensive, difficult-to-obtain experimental data.
• Traditional and machine learning methods face challenges due to specialized data label requirements, limiting practical applications.

Purpose of the Study:

• To develop a novel inverse problem formulation for discovering interpretable plasticity models.
• To leverage kinematic observations for material model identification, overcoming experimental data limitations.

Main Methods:

• Utilized a differentiable simulator with smooth constitutive updates for neural network (NN) training.
• Employed backpropagation to train NNs parameterized by kinematic observations.
• Applied digital image correlation techniques for accurate displacement measurements.

Main Results:

• Successfully inferred complex plasticity models from kinematic data.
• Demonstrated a data-efficient approach to constitutive model discovery.
• Overcame challenges associated with loading history dependence in inverse problems.

Conclusions:

• Kinematic observations can effectively identify complex material models, including plasticity.
• This method enables the generation of numerous material models, advancing fields like metamaterial design.
• The approach offers a game-changing solution for material modeling in various engineering applications.