Machine learning-based microhardness prediction and microstructural analysis of Stainless Steel 316-Inconel 718 functionally graded material through metal additive manufacturing

Purpose Functionally graded materials (FGMs) possess gradual compositional and microstructural transitions. This study aims to predict the mechanical properties of FGMs, especially the microhardness, by integrating experimentation and machine learning (ML). The work investigates the microstructures and mechanical behaviour. Design/methodology/approach The study utilized a twin-powder directed energy deposition (DED) process to fabricate a Stainless steel 316 (SS316) – Inconel 718 (IN718) based FGMs. Microstructural analysis and mechanical characterization involving uniaxial tensile, high cycle fatigue and microhardness tests are conducted. Various machine learning (ML) models are employed to predict microhardness at gradient regions. Findings The microstructural analysis revealed dendritic morphology and an increased laves phase from the gradient to the neat IN718 region. Elevated temperature uniaxial and high cycle fatigue tests revealed satisfactory yield strength and fatigue life. An increased but nonlinear trend in microhardness from SS316 to IN718 is observed. The gradient boost and random forest models successfully predicted the trends in microhardness. Practical implications The SS316–IN718-based FGMs can be used in internal combustion (IC) engine valves, where the former region provides corrosion resistance and the latter provides thermal stability at higher temperatures. Similarly, the graded forging punches can possess wear resistance at the impact surface and toughness at the core. Originality/value The FGMs are deposited uniquely in the stepover direction, which drastically reduces the anisotropy effect on properties. The study utilizes experimental values to train the ML models and predict the microhardness of gradient regions with the highest accuracy. The gradient boost and random forest models effectively capture nonlinear variations and reduce experimental runs.