Modeling and calibration of uncertainty in material properties of additively manufactured composites

Simulations quantifying the uncertainty in structural response and damage evolution require accurate representation of the randomness of the underlying material stiffness and strength behaviors. In this paper, the mean and variance descriptions of variability of strength and stiffness of additively manufactured composite specimens are augmented with random field correlation descriptors that represent the process dependence on the property heterogeneity through microstructure variations. Two correlation lengths and a rotation parameter are introduced into randomized stiffness and strength distribution fields to capture the local heterogeneities in the microstructure of Additively Manufactured (AM) composites. We formulated a simulation and Artificial Intelligence (AI)-based technique to calibrate the correlation length and rotation parameter measures from relatively few samples of experimentally obtained strain field observations using Digital Image Correlation (DIC). The neural networks used for calibrating the correlation lengths of Karhunen-Loève Expansion (KL expansion) from the DIC images are trained using simulated stiffness and strength fields that have known correlation coefficients. A virtual DIC filter is used to add the noise and artifacts from typical DIC analysis to the simulated strain fields. A Deep Neural Network (DNN), whose architecture is optimized using Efficient Neural Architecture Search (ENAS), is trained on 150,000 simulated DIC images. The trained DNN is then used for calibration of KL expansion correlation lengths for additively manufactured composite specimens. The AM composites are loaded in tension and DIC images of the strain fields are generated and presented to the DNNs, which produce the correlation coefficients for the random fields as outputs. Compared to classical optimization methods to calibrate model parameters iteratively, neural networks, once trained, efficiently and quickly predict parameters without the need for a robust simulator and optimization methods.