Data-driven prediction of air bending

Experimental data can provide a rich source of information and opportunities for the improvement of production processes. This paper provides a conceptual description of a data-driven model and discusses its possible implementation strategies. Two data-driven methods are applied to air bending, a conventional sheet metal forming process. Self-improving data-driven prediction based on an available experimental database is implemented for two calculation approaches: the gray box and the black box. A significant improvement in terms of springback prediction has been achieved with only a limited set of training data: with 100 training samples the relative error decreases from 6.75%, as given by a state-of-the-art analytical model, to 1.6%. Additionally, a material model-free approach is studied. In this model, no predefined material model is used and all necessary material related parameters are calculated based on experimental data. With only 10 training samples and using Kriging interpolation, this approach resulted in a relative error of 5%. Such results demonstrate that data-driven calculation methods allow for significant improvements in prediction accuracy with limited number of test samples. In the case of air bending, this allows the use of a data-driven response purely based on experimental data instead of on a material law that is typically expensive to obtain. (C) 2019 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/) Selection and peer-review under responsibility of the organizing committee of SHEMET 2019.