One for all: Universal material model based on minimal state-space neural networks.

Area of Science:

• Computational materials science
• Mechanical engineering
• Artificial intelligence in engineering

Background:

• Computational models are crucial for optimizing structural stiffness and strength.
• Current advanced material models are often mathematically complex and material-specific, limiting practical engineering use.
• A need exists for a universal, yet accurate, material modeling framework.

Purpose of the Study:

• To develop a novel recurrent neural network (RNN) framework for material modeling.
• To introduce "Minimal State Cells" as a core component of the RNN framework.
• To create a universal material model applicable to various engineering materials.

Main Methods:

• Development of a recurrent neural network framework incorporating Minimal State Cells.
• Application of the framework to datasets from four distinct material classes.
• Validation against experimental or simulation data for stress-strain responses.

Main Results:

• The RNN framework accurately reproduces 3D stress-strain responses for arbitrary loading paths.
• The model successfully replicates the state space of conventional material models.
• The framework demonstrates flexibility in capturing the mechanical behavior of diverse materials.

Conclusions:

• The developed RNN framework offers a universal and flexible approach to material modeling.
• Minimal State Cells enable accurate and interpretable material state representation.
• This innovation has the potential to significantly advance engineering practice in material selection and structural design.