Predicting and designing the rheology of disordered networks with geometric deep learning

Disordered semiflexible polymer networks form the structural backbone of biological cells and tissues. Their mechanical properties are highly nonlinear and sensitive to subtle structural details, making them difficult to predict without computationally expensive simulations. To address this challenge, we introduce a geometric deep learning framework that predicts mechanical behavior directly from network structure. We train equivariant graph neural networks that robustly predict the full linear and nonlinear viscoelastic response of model networks using only their initial, undeformed configurations as input. Beyond prediction, we show how this framework can be inverted to perform inverse design, efficiently generating novel network architectures with tailored mechanical properties. This data-driven approach opens a new avenue for understanding and engineering the mechanics of disordered network materials.