Transfer learning for metamodel construction to enable uncertainty quantifications in crash design based on scarce data availability

This work investigates the possibility to obtain metamodels to use in uncertainty quantification when little to no data points are available. This scenario occurs in the early-stage vehicle design for crashworthiness. As long as a suitable model of the new design is not available, the data from predecessor designs might be of aid. The aim of this study is to compute the metamodel of a crash box whose geometrical features differ partially from those of past comparable structures. To reach this goal, a transfer learning (TL) method is investigated. The finite element simulations data of the past crash boxes constitute the source domain; the few available simulations of the product of interest, instead, create the target domain. The TL method learns the basic knowledge from the source domain and then is readjusted based on the target domain. This approach constitutes an attractive starting point for the investigation of diverse types of system uncertainties. Overall, the presented method is considered especially helpful for engineers in the early-stage design process.