PREDICTION OF PROCESS-INDUCED DEFORMATIONS USING DEEP LEARNING INTERFACED FINITE ELEMENT (FE) CONSTITUTIVE MODELS

The aim of the study is to improve the predictive capacity of a Finite Element tool in relation to a rheological thermo-chemo-viscoelastic constitutive model. This enhancement specifically focuses on accurately capturing the Process Induced Deformations (PID) resulting from the polymerization of thermoset composite matrix. These deformations are due to the internal residual stress that arises from the material's inherent anisotropic properties, specifically the coefficients of thermal expansion and chemical shrinkage. The focus of the study is to accurately model the cure polymerization behaviour, which is known to have a significant impact on manufacturing defects. To account for the effect of process variables, such as maximum curing temperatures and temperature rates, a non-parametric neural network model is implemented instead of a parametric diffusion cure-kinetics model. Such model is trained using Differential Scanning Calorimetry characterization tests and is interfaced with the classical visco-elastic constitutive model to predict the evolution of thermoset resin states, which is evaluated using two cure state variables: degree of cure and glass transition temperature. This improved prediction of state transitions results in precise evaluations of internal residual stresses, leading to accurate PID predictions. Anisotropic properties of carbon/epoxy woven composite at different states of cure are used for the numerical analyses. Finally, the enhanced methodology is applied to a case study of a Z-shaped thermoset part, and the predicted PID closely associates with the experimental measures.