Evaluating the Predictability of Surface Roughness of Ti-6Al-4V Alloy from Selective Laser Melting

Herein, it is aimed in this study to assess a long-standing question, "whether reliable predictive models can be developed for surface roughness of laser additively manufactured alloys, with aggregated knowledge learned but little new experimentation?" A most comprehensive dataset is obtained from mining the literature on as-built Ti-6Al-4V alloy via selective laser melting (SLM), which includes the major SLM process parameters, average and variance of powder particle size, and resultant surface roughness. Five supervised machine-learning models and one traditional regression model are proposed to evaluate the predictability of average surface roughness Ra of SLM-produced Ti-6Al-4V alloy. Using tenfold cross validation on dependable literature data, the six models are trained and compared. To ultimately test the developed models, cubic Ti-6Al-4V samples are fabricated and the models are employed to predict their surface roughness. It is discovered that while all six models show overall satisfactory performance, the traditional regression model possesses the lowest prediction accuracy. Two of the best models, that is, Gaussian process regression and neural network models, exhibit root-mean-squared error of 0.51 and 0.58 mu m, respectively, and are championed as industry ready. Herein, an alternative path for predicting SLM contingencies is proved by this research, where resources are limited but historical experience can be mined.