Automated and Defect-free Forming of Complex Composite Parts with Machine Learning

Student thesis: Doctoral ThesisPhD


Non-crimp fabric (NCF) composites are increasingly regarded as a viable alternative to pre-impregnated composite materials. Favoured for their ability to streamline productivity, whilst offering comparable mechanical properties, these materials are first formed and subsequently infused with a resin matrix. Drape forming processes are typical in production, yet, offer little opportunity to influence deformation. Existing research in the area has focused on laboratory scale processes, with very little work looking at scalability for industrial parts. These concerns not only decrease productivity, but create inflexibility in the design and manufacture of composite components. Coupled with the continuous development of new NCFs, industry demands processes that are adaptable, scalable and purpose-built for predictive modelling.

This thesis first considers the deformation mechanics of a complex NCF, interply
shear stiffness, intraply shear stiffness and out-of-plane bending stiffness. Here, the presence of a “veil” layer revealed important characteristics that further exacerbated the material anisotropy. Subsequently, an FEA model using coupled shell and membrane elements was validated. This macroscale model allowed efficient designation of in- and out-of-plane properties via experimental characterisation.

A novel preforming process that can generate in-plane tension through discontinuous blank boundary conditions was established. Distributed Magnetic Clamping allows for highly flexible process control. This method was designed to reduce the human factor in production, and work towards an off-the-roll, intelligent resin transfer system. An efficient two-stage optimisation routine deployed a Gaussian process model for preform deformation, and Bayesian optimisation to find the optimal clamping locations. Results demonstrate that surrogate modelling is viable for magnetic distributed clamping. This lays the foundation for the development of large, translatable data sets that can maximise the impact of statistical tools.

Finally, productivity requirements motivate industrialists to form multiple plies in one stroke. Therefore, DIMAC was employed to experimentally investigate the feasibility of multiple-ply preforms. An increased number of biaxial plies formed in one stroke was possible through DIMAC, after using the key insights developed through predictive modelling and single-ply experiments. Similarly, the process was able to influence individual ply deformation for stacks with diverse fibre orientations toward the goal of high-rate manufacture.
Date of Award2 Nov 2022
Original languageEnglish
Awarding Institution
  • University of Bath
SponsorsGKN Aerospace
SupervisorEvros Loukaides (Supervisor), Vangelis Evangelou (Supervisor) & Richard Butler (Supervisor)

Cite this