Predictability assessment of as-built hardness of Ti-6Al-4V alloy fabricated via laser powder bed fusion

Evaluating the predictability of selective laser melting (SLM) process has been a persistent endeavor in the manufacturing community since the technique's inception. The ability to accurately predict the as-built hardness of fabricated parts is critical to judge its serviceability for specific applications. This investigation aims at achieving such predictability with as little as experimentation possible. In this research, reliable historical SLM data was mined on as-built Ti-6Al-4V alloy, from technical articles published over the past fifteen years. Three predictive models of Gaussian Process Regression (GPR), Neural Network (NN) and parametric Multiple Linear Regression (MLR) were built for as-fabricated bulk hardness of Ti-6Al-4V alloy manufactured via SLM. Ten-fold cross-validation training of the models involved a substantial 1,284 spatial datapoints of major SLM process parameters such as laser power, scanning speed, hatch spacing, layer thickness and volumetric energy density. Ultimate assessment of hardness predictability was performed by making predictions on ten cubic Ti-6Al-4V test coupons fabricated using a SLM machine. The GPR model stood out amongst other models, yielding a mean absolute prediction error of 6.12 HV over all samples. The overall predictability of these models came out to be in declining order of GPR > NN > MLR. Rudimentary characterization of test coupons revealed variation in observed hardness. Two tiers of robust validations, comprehensive dataset, statistical insights, actual experimental validation and material characterization make this study extremely unique for as-fabricated hardness of SLM-ed Ti-6Al-4V alloy.