Physics-informed machine learning model for computational fracture of quasi-brittle materials without labelled data

Only a few studies have focused on the simulation of the crack propagation of quasi-brittle materials under complex loading inspired by physical-informed neural networks (PINN). Guided by the energy minimization principle rather than labelled data, we reconstruct the solution of displacement field after damage to predict crack propagation using PINN which maintains the thermodynamics consistency inherited from our proposed variable four-parameter damage model. Additionally, the framework of incremental pattern performs relatively efficiently with transfer learning. Consequently, a novel method for better convergence based on domain decomposition theory is proposed to identify complex boundaries. Based on gradient pathology, we develop a finite basis algorithm to solve the ill-condition problem. Whether under uniaxial tension, pure shear or mixed mode loading, the prediction results of displacement field fit well with simulations in the literature. Our research is meaningful for improving the generalization of neural networks and accelerating the optimization process which are necessary for further engineering applications.

Automated summary

This study focuses on using physically-informed neural networks to simulate the crack propagation of quasi-brittle materials under complex loading. By minimizing energy, we use neural networks to predict crack propagation while maintaining thermodynamic consistency. The results show that our proposed method accurately predicts displacement fields under different loading conditions.