ENHANCING TRANSFER LEARNING FOR CRASHWORTHINESS STUDIES UNDER LOW DATA AVAILABILITY THROUGH SPHERE PROJECTION

This work proposes a transfer learning (TL) approach for crashworthiness analyses. It investigates the possibility to exploit data and geometrical information coming from past studies when only a few data points are available for the current product of interest. This situation occurs early on in the development, when assessing the crash safety in the automotive industry. As the final characteristics of the product under investigation are not yet entirely defined, TL can be of aid. By transferring the knowledge from past simulations or uncertainty propagation studies — the so-called source domain — to forthcoming cases — the target domain, TL can offer valuable predictions. The data coming from past similar products, however, often belong to different geometrical designs with respect to the one of interest. The TL network needs to be able to distinguish them. Therefore, the methodology can be enhanced by coupling the TL with an already existing geometry classification technique: sphere projection. In this way, the TL approach becomes able to predict the new structural behavior under uncertainties. We demonstrate this by applying TL to an explicatory bonnet example. The obtained results picture the enhanced TL approach as an attractive starting point. With respect to past studies, the proposed approach gives more flexibility to TL for crash analyses. It can now help to enhance UQ studies for crashworthiness when only a few data are available.