Deep learning predicts path-dependent plasticity.

Area of Science:

• Computational mechanics
• Materials science
• Machine learning

Background:

• Plasticity theory describes material yield loci and work hardening under deformation.
• Current models rely on simplifications like yield criteria, flow rules, and work equivalence, which struggle with complex material histories and microstructures.
• Iterative algorithms like return mapping are needed to trace yield surface evolution.

Purpose of the Study:

• To investigate if sequence learning with recurrent neural networks can overcome limitations in traditional plasticity theory.
• To explore an alternative approach for developing constitutive models that account for material history and microstructure.

Main Methods:

• Utilized recurrent neural networks (RNNs) for sequence learning.
• Applied deep learning techniques to model material behavior.
• Demonstrated that standard plasticity assumptions are not required within the RNN framework.

Main Results:

• Recurrent neural networks effectively learn history- and microstructure-dependent material behaviors.
• The proposed deep learning approach bypasses the need for complex yield criteria, flow rules, and iterative algorithms.
• Established foundations for novel constitutive models through deep learning.

Conclusions:

• Deep learning, specifically RNNs, provides a powerful alternative to established plasticity formulations.
• This approach simplifies the modeling of complex material responses by learning directly from data.
• Paves the way for more accurate and efficient material modeling in engineering applications.