Self-consistent recurrent neural network for path dependent deformation.

Area of Science:

• Computational Mechanics
• Machine Learning
• Materials Science

Background:

• Machine learning (ML) models are increasingly used as surrogates in engineering, but path-dependent problems pose challenges.
• Generic ML models may not inherently satisfy temporal constraints crucial for physical accuracy in simulations.

Purpose of the Study:

• To examine deformation path characteristics from finite element simulations.
• To identify key requirements for physically meaningful surrogate modeling in path-dependent problems.
• To propose a novel Recurrent Neural Network (RNN) approach addressing these requirements.

Main Methods:

• Analysis of deformation paths from finite element simulations.
• Identification and formalization of the truncation condition (future inputs not influencing past outputs).
• Development of a customized RNN transition function enforcing truncation and consistency across time discretizations.

Main Results:

• The proposed RNN transition function effectively enforces the truncation condition.
• The model ensures consistency across different temporal resolutions in deformation history modeling.
• Demonstrated improvement in physically consistent damage initiation estimation.

Conclusions:

• The customized RNN approach enhances the reliability of surrogate models in computational mechanics.
• Enforcing physical properties like truncation and consistency is vital for accurate ML-based engineering predictions.
• This work advances the development of robust surrogate models for complex physical processes.