Dissertation Defence: Data-Driven Processing–Property–Defect Relationships in the Forming of Flax Fibre Composite Reinforcements

Olivia Helena Margoto, supervised by Drs. Abbas Milani & Yasmine Abdin, will defend their dissertation titled “Data-Driven Processing–Property–Defect Relationships in the Forming of Flax Fibre Composite Reinforcements” in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Mechanical Engineering.

An abstract for Olivia Helena Margoto’s dissertation is included below.

Examinations are open to all members of the campus community as well as the general public. This examination will be offered in hybrid format.  Registration is not required to attend in person; however, please email abbas.milani@ubc.ca to receive the Zoom link for this exam.

Abstract

Natural Fibre-Reinforced Composites (NFRCs) already offer lightweight and sustainable alternatives for several industrial applications including those in automotive sector. However, similar to synthetic composite options, forming-induced defects such as wrinkles can reduce targeted performance of NFRCs. Predicting these defects is particularly challenging for natural fibre woven reinforcements due to the inherent variability in fibre geometry and resulting fabric properties.

This study applies a data-driven approach to predict wrinkle formation in flax (2×2 twill as well as biaxial non-crimp) and glass (as synthetic benchmark) twill weave reinforcements. The fabrics’ shear, bending, tensile, and friction behaviors were experimentally characterised to capture forming-relevant mechanical properties. The fabrics were subsequently formed in multi-layer configurations over a tetrahedron and square tools, followed by 3D scanning to quantify wrinkle distributions for each material alternative. Unsupervised data-driven methods were next used to assist multivariate analysis and identification of potential post-forming processing–property–defect relationships. Two types of descriptors were calculated from forming experiments to quantify wrinkle severity: i. structural response metrics, representing Surface Distortion (SD), Volume Distortion (VD), a Wrinkle Index (WI), and Compliance (CP); and ii. image texture descriptors, where post-forming surface deformations were transformed into grayscale maps from which Haralick texture features, including contrast, dissimilarity, energy, and homogeneity, were extracted. Combined with the fabric design parameters such as weave type, forming orientation, number of layers, grammage, and thickness, those descriptors were used to train i. Linear Regression (LR), ii. Partial Least Squares (PLS) regression, and iii. Artificial Neural Network (ANN) models. The models were evaluated using both k-fold and Monte Carlo cross-validation strategies in their ability to predict the described wrinkle severity descriptors. Results clearly showed that flax twill reinforcements exhibited the best forming performance, characterised by the lowest overall occurrence of out-of-plane forming-induced defects, followed by the biaxial fabrics, while the glass twill weave exhibited the poorest performance under both the tetrahedron and the square forming tools. For the latter geometry, a 45° fabric orientation was found to be optimal for all woven reinforcements, as it maximised in-plane shear accommodation and minimised wrinkle formation. This effect was most pronounced for the flax twill with 367 gsm areal weight, which demonstrated the best overall formability due to its favourable balance between bending stiffness and shear compliance. In contrast, the biaxial non-crimp fabric exhibited limited orientation sensitivity, as their stitched architecture restricted yarn rotation and shear deformation. Linear models, LR and PLS, exhibited similar performance, accurately predicting defect severity through SD, dissimilarity, and homogeneity, with R2 ~ 0.70. The ANN models were able to capture nonlinear relationships between the input material properties and the wrinkle severity descriptors, achieving a superior performance R2 ~ 0.90 for SD, VD, and dissimilarity. In addition to achieving the best and most consistent predictive performance across all descriptors and models, the dissimilarity feature effectively described both defect severity and fabric architecture effects, with lower values for glass twill reinforcements and higher values for the flax non-crimp fabric reinforcement. Results also revealed that the effective bending stiffness, shear stiffness, and fabric orientation are among the most influential material factors in forming-induced defect severity, as confirmed by SHAP analysis. Overall, the findings demonstrated that integrating structural/image-texture descriptors with process parameters and mechanical properties provide a promising machine-learning framework for rapidly predicting the forming quality of NFRCs.