Directed Gaussian process metamodeling with improved firefly algorithm (iFA) for composite manufacturing uncertainty propagation analysis

A computationally effective and physically accurate metamodeling approach is demonstrated to analyze, under uncertainties, the spring-in angle deformation for composite manufacturing processes. Various uncertainties are inevitably present in this manufacturing process due to the heterogeneous thermo-mechanical properties of the composite materials. Analysis of uncertainty propagation using the direct Monte Carlo approach is computationally prohibitive, which calls for the employment of machine learning techniques and surrogate models or metamodels such as Gaussian processes (GP). While these approaches are promising, tuning model parameters and optimizing the hyperparameters are essential to predictive modeling performance. So far, most existing approaches rely on empirical experience through trial and error. Randomly selecting these hyperparameters results in excessive computational cost and poor convergence results. A nature-inspired methodology has been developed to guide the GP in selecting and optimizing the hyperparameters for the uncertainty propagation analysis of composite manufacturing processes. An improved firefly algorithm (iFA) takes account of the environmental factor. It disregards the contribution of a constant attractiveness factor, which in turn accelerates the convergence rate at the early stages of the generation and boosts the immunity of the proposed algorithm. The proposed methodology enabled selection of the proper combination of the factors for the GP and showed its merits over other state-of-the-art deterministic/metaheuristic algorithms, which is further confirmed by various nonparametric, multiple comparison tests.