This study presents an imperfection-tolerant, surrogate-assisted framework for the multi-objective optimisation of variable-stiffness (VS) composite cylinders that explicitly incorporates experimentally measured geometric imperfections. Principal component analysis (PCA) is applied to extract dominant imperfection modes from experimental data, and Latin hypercube sampling (LHS) is used to generate statistically consistent synthetic fields, which are subsequently mapped onto nonlinear finite element (FE) models. Linear buckling and geometrically nonlinear collapse analyses are performed under axial compression to determine the ideal and actual load-carrying capacities, from which the knockdown factor (KDF), quantifying imperfection sensitivity, is derived. Gaussian Process Regression (GPR) surrogates are trained to predict the mass and collapse loads of perfect and imperfect geometries with high cross-validated accuracy, while KDF is computed as their ratio. The framework enables simultaneous optimisation of three objectives: mass minimisation, collapse-load maximisation, and KDF maximisation by using Bayesian Optimisation (BO) and the Non-dominated Sorting Genetic Algorithm II (NSGA-II) independently. Results demonstrate that integrating experimentally informed imperfections with surrogate-based optimisation captures the key physical trends governing buckling and imperfection sensitivity, while achieving substantial computational savings relative to direct nonlinear analyses, and that both optimisers yield consistent Pareto fronts featuring smooth, manufacturable fibre trajectories that balance lightweight efficiency, strength, and robustness.

This study presents a Bayesian Optimisation (BO) framework for the mass minimisation of variable-stiffness (VS) composite cylinders under multiple buckling constraints, incorporating manufacturing limitations derived from filament winding processes. A computationally efficient single-curvature finite element model is used to evaluate the linear buckling response of multilayered shells. BO simultaneously optimises fibre paths, number of layers, and thickness distribution, achieving comparable or improved performance relative to a Genetic Algorithm (GA) while reducing simulation time by up to 70 %. Across most design loads, BO delivers structurally efficient solutions with smooth thickness transitions and local stiffness tailoring. Although GA outperformed BO in the highest load case in terms of weight and buckling capacity, BO retained competitive performance and demonstrated higher modal richness. Buckling mode analyses revealed that BO designs support mixed-mode instabilities with greater circumferential complexity, enhancing structural adaptability. In contrast, GA designs exhibited more uniform fibre paths and axial-dominated modes, reflecting conservative reinforcement strategies. These findings highlight the capability of BO to exploit complex design spaces more effectively, offering a scalable and data-efficient alternative to traditional optimisation methods. The proposed framework is particularly well suited for high-fidelity, simulation-driven design of advanced composite structures where computational cost and manufacturability are critical constraints.