Influence of tensile properties on hole expansion ratio investigated using a generative adversarial imputation network with explainable artificial intelligence

Abstract

Hole expansion ratio is widely used to estimate the stretch flangeability of sheet metals which is a critical property of formability and to evaluate the efficiency of a forming process. Although many experiments were conducted in the past to identify the key tensile properties affecting hole expansion ratio, their results failed due to the data scarcity. This work demonstrates a machine learning framework coupled with imputation methods to augment both the quantity and quality of collected experimental data. Especially, a generative adversarial imputation network (GAIN) is used to impute the missing tensile properties in the collected experimental data. With the imputed data, the hole expansion ratio is predicted through an extra tree regressor. In terms of the imputation performance, GAIN resulted in the lowest root mean square error of 0.09146 when 50 known tensile properties are randomly removed and imputed with GAIN. In terms of the hole expansion ratio prediction performance, the extra tree regressor showed the lowest root mean square error of 0.124 compared to other machine learning models. Finally, the influence of each tensile property on the hole expansion ratio is analyzed using Shapley additive explanations, an explainable artificial intelligence technique. In this study, the influences of various tensile properties on hole expansion ratio were quantitatively determined for the first time via machine learning and this analysis will accelerate the exploration of sheet metals with high formability performances.