Neural network-assisted sensitivity analysis applied to finite-element simulations of plant-fiber microdroplet tests

Abstract

Microdroplet tests are widely used for interface characterization in fiber-reinforced composites. However, their interpretation remains challenging due to complex, non-uniform stress states and the interplay of multiple geometric and mechanical factors. These challenges are further amplified in biocomposites, where natural fibers introduce additional variability in shape, orientation, and adhesion quality. This study presents a surrogate-based methodology to systematically quantify these sensitivities. A fully parameterized finite element model captures the geometric and interfacial complexities of natural fiber microdroplet tests. To efficiently explore parameter dependencies, a surrogate model based on artificial neural networks (ANNs) is trained on a comprehensive dataset of finite element simulations. Sensitivity analyses reveal strong effects of geometrical variability and mixed-mode behavior, questioning the validity of the commonly used scalar strength metric—interfacial shear strength (IFSS). By leveraging first- and second-order sensitivity analyses, we demonstrate how non-linear parameter interactions shape the macroscopic stress–displacement curves. These findings open up the possibility of using the surrogate model for inverse identification of interfacial parameters for use in composite-scale models, reducing reliance on repeated finite element simulations that often accompany interface characterization experiments.