Data Driven Modeling of Interfacial Traction Separation Relations using Thermodynamic Consistent Neural Network (TCNN)

For multilayer structures in thin substrate systems, the interfacial failure is one of the most important reliability issues. The traction-separation relations (TSR) along fracture interface, which is often a complicated mixed-mode problem, is usually adopted as a representative of the adhesive interactions of a biomaterial system. However, the existing theoretical models lack complexity and are not able to fit with real-world TSRs obtained with end loaded split beam (ELS) experiments. The neural network fits well with the experimental data along the loading paths but fails to obey physical laws for area not covered by the training data sets, due to the lack of mechanics in pure neural network fitting with training data sets. In this paper, a thermodynamic consistent neural network (TCNN) is established to model the interface TSRs with sparse training data sets. Three thermodynamic consistent conditions are considered and implemented with neural network model. By treating these thermodynamic consistent conditions as constraints and implementing as loss function terms, the whole traction surface is constrained to provide reasonable results. The feasibility of this approach is approved by comparing the modeling results with different number of physical constraints. Moreover, the Bayesian optimization algorithm is adopted to optimize the weighting factors of the TCNN to overcome the convergence issue when multiple constraints are in present. The numerical implementation results demonstrated well behaved prediction of mixed-mode traction separation surfaces in terms of high agreement with experimental data and damage mechanics contained thermodynamic consistencies. The proposed approach opens doors to a new autonomous, point-to-point constitutive modeling concept for interface mechanics.