Bayesian optimisation of resolved interlaminar stresses in composite laminates using high-aspect-ratio discontinuous Galerkin finite elements

Abstract

Laminate optimisation is a key route to improved composite structures, linking material choice, fibre orientation and stacking sequence to structural performance and manufacturability. Using this design freedom to optimise local interlaminar-stress severity directly is much harder, since the governing stress field is three-dimensional, highly localised and costly to resolve accurately. This paper presents a coupled discontinuous Galerkin (DG) finite-element and Bayesian optimisation (BO) framework for treating local interlaminar-stress severity as the optimisation objective itself. Candidate laminates are chosen from a finite material-and-fibre-angle library and evaluated with a full 3D DG finite-element solver, whose high-aspect-ratio meshes are central to the efficiency of the optimisation framework: they concentrate refinement around monitored stress regions and resolve the local three-dimensional stress fields without requiring fine elements throughout the surrounding domain. Scalar patch objectives are constructed from the resolved interlaminar normal and shear stresses, including hardmax, smoothmax and monotonic power/log transforms, to test how objective representation affects BO performance without changing the underlying candidate laminate set. Benchmarks are straight free-edge and corner-stressed rectangular-cutout laminates. Against a matched genetic-algorithm (GA) baseline, BO consistently gives higher exact-optimum reliability and better final ranks. In the free-edge benchmark, the best transformed objectives identify the optimum in every run; GA needs 7 times more evaluations and 6.7 times longer mean time to identify the same optimum. In the cutout benchmark, BO finds the optimum after sampling only about 1.2% of the feasible space; GA needs 8 times more evaluations and 7.5 times longer mean time to do so.