Convolutional neural networks for additively manufactured hot work steel H11: classification and prediction of mechanical properties

Abstract The production of parts with high strength and consistent quality through metal additive manufacturing is a challenging task. Machine learning methods may help optimize the production process or improve materials. In this work, convolutional neural networks (CNNs) are developed for the classification of hot work steels and for the prediction of mechanical properties of these steels from microstructure images recorded by scanning electron microscopy (SEM). A total of 18 samples of two different hot work steels fabricated by laser-based powder bed fusion (PBF-LB/M) are investigated, each with or without preheating and with one of five different post-heat treatments. The steels are standard steel H11 and a newly developed modification intended for the same application. The goal of the heat treatments and the modification of the steel is to achieve comparable crack-free properties without the need for a preheating process. The microstructure images of the 18 samples are accurately distinguished by a classification CNN with an overall accuracy of 98%. Furthermore, regression CNNs for the prediction of yield strength (YS) and ultimate tensile strength (UTS) generalize well and predict YS and UTS for new untrained samples with average relative errors of only 4.1% and 3.9%, respectively.