AI/ML for the Quantification of Process-Induced Uncertainty in Additively Manufactured Composites

Variability in properties of heterogeneous composites arises from manufacturing processes, service loads, or environmental factors and significantly impacts the macroscopic structural performance. Uncertainty quantification (UQ) offers a framework for measuring microstructure and property uncertainties and evaluating their effects on structural performance, thus enhancing confidence in simulation outcomes. Comprehensive UQ can facilitate certification by simulation for composite materials and manufacturing processes. Extensive experimentation now needed for the certification of aerostructures can be reduced with validated UQ methods. This paper explores strategies for using artificial intelligence/machine learning (AI/ML) methods for uncertainty quantification in additively manufactured (AM) composite parts. We present two examples of applying AI/ML strategies to quantify property and strain variability in parts reinforced with continuous fibers using a Fused Filament Fabrication printer. First, a stochastic model that represents the spatial uncertainty of properties in the printed composites was formulated. The parameters of the spatial uncertainty model were calibrated from limited experimentally observed and denoised strain maps using a trained convolutional neural network. Second, we demonstrated that a decoder-only transformer neural network can be trained to serve as a microstructure-aware constitutive model that can predict a composite’s homogenized (macro-scale) transverse properties.