Experimentally trained physic-informed neural network as material model

The authors propose a new approach for the data-driven discovery of composite material models leveraging on physic informed mechanistic neural networks. The methodology is unsupervised since the surrogate neural network can learn the unexplicit constitutive law from the strain field and global force data of mechanical tests, that can be easily collected with digital image correlation technique. The approach integrates the distinctive characteristics of both mechanistic data science and physic informed neural networks: the neural network architecture is designed to predict a constitutive law respecting the material symmetry and decoupling of the axial and shear response; the data-driven model is trained with a custom loss function enforcing the equilibrium constraints between external and internal energy. The results of the training is a physic-informed neural network predicting the response of composite. Using experimental data on tensile tests of carbon fiber woven reinforced epoxy specimens, authors demonstrate the capability of the data-driven method to efficiently discover the mechanical response of composite material with a reduced set of experiments.