A strategy to formulate data-driven constitutive models from random multiaxial experiments.

Area of Science:

• Materials Science
• Computational Mechanics
• Machine Learning

Background:

• Accurate constitutive models are crucial for predicting material behavior under complex loading conditions.
• Traditional models may struggle with the intricate responses of elastic-plastic materials subjected to combined stresses.
• Data-driven approaches offer a promising alternative for capturing complex material behaviors.

Purpose of the Study:

• To develop and validate a test technique and computational framework for data-driven surrogate constitutive models.
• To capture the in-plane stress response of isotropic, elastic-plastic materials under combined normal and shear stresses.
• To assess the accuracy and feasibility of using feed-forward neural networks for this purpose.

Main Methods:

• A computational framework utilizing feed-forward neural networks (NNs) was developed.
• Virtual experiments (Finite Element simulations) were conducted on a thin-walled specimen under random axial displacement and rotation histories.
• The material was modeled as an isotropic, rate-independent elastic-plastic solid obeying J2 plasticity with isotropic hardening.

Main Results:

• A training dataset was assembled from the virtual experiments.
• Two distinct surrogate models based on neural networks were trained and evaluated.
• Both surrogate models demonstrated effectiveness and comparable accuracy in predicting material response to random multiaxial strain histories.

Conclusions:

• The proposed data-driven approach using neural networks is feasible for creating surrogate constitutive models.
• The developed framework accurately captures the complex response of elastic-plastic materials under combined in-plane stresses.
• This technique provides a valuable tool for enhancing the accuracy of engineering simulations involving material behavior.