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

Abstract

Disordered fibrous networks play vital mechanical roles but are difficult to model due to their sparse connectivity and complex nonaffine deformation. We introduce a physics-informed, graph-learning-based Network Mechanics Prediction (GNMP) that predicts deformation and stress-strain responses of 2D semiflexible networks with reliable accuracy and >10× efficiency gains over molecular dynamics. GNMP integrates graph-attention message passing, multiscale physical embeddings, and a bond-length-guided scheduler to capture key nonaffine rearrangements. Experiments on Poisson-ratio-aligned 3D-printed networks show consistent deformation motifs, descriptor trends, and comparable log-area changes (∼0.46 vs ∼0.24) while reducing the geometry-inference time to ∼0.13 s per step. GNMP offers a generalizable route for rapid, topology-aware mechanical prediction in fibrous network materials and supports the accelerated design of biomimetic soft tissues, flexible conductors, and related network-based systems.