Surrogate modeling strategy for random fields with uncertain correlation lengths

Résumé fourni par la source

Thin-walled shell structures achieve high load-bearing capacities with low material consumption. However, they are highly susceptible to stability issues caused by geometric imperfections. To account for these imperfections, random fields are incorporated into finite element (FE) models, and the structural response is evaluated with Monte Carlo simulations. Time-consuming FE analyses can be replaced by an artificial neural network (ANN) surrogate model. When applying the Karhunen-Loève expansion (KLE), the shape of the random field is determined by a correlation length, which may be affected by epistemic uncertainty. Consequently, the random field is modeled using fuzzy correlation lengths, leading to a further increase in computation time. To reduce this effort, an interpolation-based surrogate modeling strategy is proposed and compared with the conventional approach of solving the KLE eigenvalue problem. Within a multilevel surrogate modeling framework, the ANN is combined with least-squares polynomials to perform a structural optimization including fuzzy stochastic analysis. The introduced model is then applied to a fiber composite plate to optimize its thickness and fiber orientation angle. Therefore, both single-objective constrained and multi-objective unconstrained optimization are conducted.