Predicting the mechanical properties of biopolymer gels using neural networks trained on discrete fiber network data

Biopolymer gels, such as those made out of fibrin or collagen, are widely used in tissue engineering applications and biomedical research. Moreover, naturally assembles into gels vivo during wound healing thrombus formation. Macroscale biopolymer gel mechanics dictated by the microscale fiber network. Hence, accurate description can be achieved using representative volume elements (RVE) that explicitly model discrete networks microscale. These RVE models, however, cannot efficiently to macroscale due challenges computational demands multiscale coupling. Here, we propose use an artificial, fully connected neural network (FCNN) capture behavior models. The FCNN was trained on 1100 subjected 121 biaxial deformations. stress data from RVE, together with total energy condition incompressibility surrounding matrix, were determine derivatives unknown strain function respect deformation invariants. During training, loss modified ensure convexity symmetry its Hessian. A general coded a user material subroutine (UMAT) software Abaqus. In this work, finite element simulations our UMAT. We anticipate work will enable further integration machine learning tools mechanics. It also improve modeling biological materials characterized structure.