Physics-Informed Graph Neural Networks for Predicting Deformation in Disordered Fibrous Materials.

Area of Science:

• Materials Science
• Computational Mechanics
• Machine Learning

Background:

• Disordered fibrous networks are crucial in many applications but challenging to model due to complex deformations.
• Existing methods struggle with sparse connectivity and nonaffine rearrangements in these networks.

Purpose of the Study:

• To introduce a novel physics-informed, graph-learning approach for predicting the mechanical responses of 2D semiflexible networks.
• To achieve high accuracy and significant efficiency gains compared to traditional simulation techniques.

Main Methods:

• Developed Network Mechanics Prediction (GNMP), a graph-learning model integrating graph-attention message passing and multiscale physical embeddings.
• Incorporated a bond-length-guided scheduler to effectively capture nonaffine rearrangements.
• Validated GNMP using experiments on 3D-printed networks with controlled Poisson ratios.

Main Results:

• GNMP demonstrated reliable accuracy in predicting deformation and stress-strain responses.
• Achieved over 10x efficiency improvement compared to molecular dynamics simulations.
• Experimental validation showed consistent deformation patterns and reduced geometry-inference time.

Conclusions:

• GNMP provides a generalizable framework for rapid, topology-aware mechanical predictions in fibrous materials.
• This approach accelerates the design of biomimetic soft tissues, flexible conductors, and other network-based systems.