A data-driven approach for imperfection-insensitive thin-shell structure design via localized dimple imperfections and gradient boosting

Thin-shell structures exhibit an unpredictable buckling behavior caused by their extreme sensitivity to localized imperfections. This work presents a data-driven framework to obtain imperfection-insensitive thin-shell structures based on an approach that replaces traditional eigenmode-based imperfection modeling with localized dimple imperfections. While the framework is general in nature, the study focuses on one particular kind of thin-shell structure, the Collapsible Tubular Mast (CTM), increasingly used in ultralight deployable space structures. When deployed, the structures experience a bending loading which can result in the boom buckling. A skew-normal distribution is shown to describe the distribution of resulting buckling moment when imperfections are seeded in the initial boom geometry, leading to the adoption of Natural Gradient Boosting (NGBoost) for probabilistic predictions under two bending directions. The models estimate mean and standard deviation of the buckling moment for varying boom design parameters, thereby capturing heteroscedastic uncertainty arising from geometric imperfections. Multi-Objective Optimization (MOO) techniques then integrate the predictive models to balance competing objectives, maximizing average buckling capacity while minimizing its variability. Results reveal distinct pareto-optimal designs that can achieve high buckling loads with reduced imperfection-sensitivity. This framework highlights the importance of local imperfection modeling and probabilistic data-driven methods in advancing robust thin shell design for next-generation deployable space systems.