Low-cycle fatigue of laser powder bed fusion-processed AlSi10Mg using recycled powder: Experiments and machine learning-assisted lifetime prediction

As additive manufacturing technologies advance, the increased use of recycled powder feedstock becomes inevitable. However, recycling may compromise the purity and quality of material inputs, potentially leading to inferior component properties. In this study, strain-controlled Low-Cycle Fatigue (LCF) tests were performed on laser powder bed fusion-processed AlSi10Mg using recycled powder. The use of recycled powder led to an increased oxygen content, resulting in more pores in the microstructure. The LCF tests covered various strain amplitudes under tension-compression for both horizontally and vertically built specimens. After the initial softening, the cyclic response stabilized, with the Hall-Petch effect identified as the main strengthening mechanism due to the eutectic cell walls, regardless of build direction. Investigations into the damage mechanisms revealed deposition defects as the main factor influencing transgranular crack initiation and propagation. Horizontally built specimens exhibited shorter fatigue lifetimes due to a higher number of deposition defects apparently caused by their positions on the build platform. A physics-informed neural network, combined with a strain-life approach, was proposed to predict the fatigue lifetime of small datasets and account for the damaging effects of deposition-related defects. The predicted data showed a good correlation with the experimental results. • Powder recycling increased the oxygen content, leading to more pores in the microstructure. • The specimen position on the build platform had a greater impact than powder recycling on the severity of deposition defects. • Transgranular cracking, linked through the defects, was identified as the main failure mode under low-cycle fatigue loading. • The Hall-Petch effect, attributed to the eutectic cell walls, was identified as the primary strengthening factor. • A physics-informed neural network was proposed to predict fatigue life of small datasets and account for multiple defects.