Impact of Sampling Strategy on the Accuracy of Surrogate Models for Structural Analysis

In modelling and simulation, approximations, or surrogate models, or metamodels, are commonly employed to approximate outcomes of complex simulations that are computationally intensive to evaluate. Surrogate models are used to map the design variables (inputs) to response variables (outputs) when the actual relationship between the two is unknown, not well understood or computationally expensive to evaluate [1]. Various approximation techniques like response surface models, Kriging, radial basis functions, neural networks, etc. are utilized for surrogate modelling. These methods inherently rely on using a predefined dataset that forms the basis of training, validation and predictive capabilities of these surrogate models. Thus, the accuracy of these surrogate models is directly related to the quality of the underlying dataset used. Design of Experiments (DOE) methods are typically used to generate these datasets. The quality of the datasets is influenced by the choice of the DOE methods that are employed as each DOE method uses a different strategy to sample the design space. Various sampling strategies like full factorial, fractional factorial, Latin hypercube sampling (LHS), Sobol sampling, etc. are available. [2] provides a state-of-the-art review on different approximation and DOE techniques available. This study examines the impact of 3 sampling techniques- full factorial, fractional factorial, and Sobol sequencing- on the accuracy and efficiency of surrogate models created using radial basis functions (RBF) for a structural analysis use-case.