A physics-guided graph neural framework for real-time reconstruction of mechanical microenvironments in engineered cartilage scaffolds

Abstract The spatially heterogeneous mechanical microenvironment within engineered cartilage scaffolds critically regulates cell fate and extracellular matrix maturation, yet its real-time characterization remains limited by the computational burden of high-fidelity finite element analysis (FEA) simulations. This study develops Phys-MRGNet, a topology-conditioned physics-guided graph neural network that rapidly reconstructs the mechanical microenvironment of a 3D-printed cartilage scaffold under varying material states and loading conditions. The network was trained on finite-element simulation sequences spanning an experimentally constrained viscoelastic parameter domain of the scaffold, with short, medium, and long edges modeled as distinct mechanical relations and a graph-Laplacian-based mechanical field consistency regularizer without extra learnable parameters. On 20 independently held-out FEA sequences comprising 4.55 × 10 6 node-step predictions, the model achieved coefficients of determination of 0.9806 and 0.9805 for interior and domain-boundary constitutive states, and a 9% lower hotspot-edge deviation than the variant without the physical term. Inference takes approximately 0.1 s per sequence, roughly 1.2 × 10 3 times faster than batch FEA on our hardware. Phys-MRGNet reproduces finite-element strain fields with high accuracy and enables rapid screening of mechanical conditioning strategies for engineered cartilage maturation through a Unity3D digital-proxy prototype, providing a mechanically relevant field descriptor for mechanobiological analysis of scaffold culture.