Metamodel Sensitivity to Sequential Adaptive Sampling in Crashworthiness Design

A study is conducted to determine the sensitivity of 2 topologically distinct metamodel types to variations in the experimental design brought about by sequential adaptive sampling strategies. The study focuses on examples encountered in crashworthiness design. Three sampling strategies are considered for updating the experimental designs, namely (i) a single stage approach, (ii) a sequential approach and (iii) a sequential domain reduction approach with higher densities in local regions. The experimental design type is the Space Filling Method based on maximizing the minimum distance between any two design points within a subdomain. Feedforward Neural Networks (NN) and Radial Basis Function Networks (RBF) are compared with respect to their sensitivity when applied to these strategies. A large set of independent checkpoints, constructed using a Latin Hypercube Sampling method is used to evaluate the accuracy of the various strategies. Five examples are used in the evaluation, namely (i) simple two-variable two-bar truss, (ii) the 21 variable Svanberg problem, (iii) a 7 variable full vehicle crash example, (iv) a 11 variable knee impact crash example and (v) a 5 variable head impact example. The examples reveal two main characteristics, namely that, while expensive to construct, NN committees tend to be superior in predictability whereas RBF networks, although much cheaper to construct can, in some cases, be highly sensitive to irregularity of experimental designs caused by subdomain updating. However, this conclusion cannot be extended to the three crash problems tested, since the RBF networks performed consistently well for these examples.