Automated fiber placement introduces defects such as gaps and overlaps during production, which add uncertainty to the mechanical behavior of composite parts. This study describes a probabilistic methodology which allows the direct introduction of random defects to the composite part and the calculation of the allowable design values through probabilistic virtual testing. The method is applied on a stiffened panel and is based on measurements regarding the random defect magnitude. An offline database of knock-down factors for stiffness and strength properties is first generated via detailed defected FE models. The factors are then assigned to the panel FE model accordingly, and via a Monte Carlo procedure the statistical response sample of the part is calculated. The process is further accelerated by a machine learning technique able to emulate the relationship between the random input and the selected responses. Several parametric analyses reveal that the number of defects on the panel is often more sensitive than the defect size, while the intensity of the defect at the through-the-thickness direction and the affected plies have a rather strong effect. Moreover, the conservatism of the deterministic design approach based on safety factors is proven by a direct comparison to the described probabilistic approach.
