Deep Learning-Based Prediction of Mechanical Properties From Metallographic Micrographs Toward a Circular Aluminum Society

Abstract

To develop a circular aluminum society, optimizing the properties of recycled aluminum is an important issue, but it requires considerable effort to develop new alloys and evaluate their properties. In this study, a deep-learning method for predicting the mechanical properties of alloys was developed. Mechanical properties of aluminum alloys are known to be affected by constituent phases and voids in the material. However, in this method, mechanical properties can be accurately predicted from micrographs without requiring information on constituent phases or density, which are typically obtained from XRD (x-ray diffractometers) data and density measurements. Using Grad-CAM (gradient-weighted class activation mapping) for verification, we confirmed that constituent phases and voids can be identified through contrast differences, and that their different effects on mechanical properties were properly captured by the model.