In this study, the influence of selective laser melting (SLM) parameters on the surface characteristics of “single-layer” raster scanning deposits, with Ti6Al4V powder, is thoroughly investigated through experimental, computational, and data-driven approaches. In layer-wise manufacturing processes, such as SLM, the evaluation and understanding of the surface characteristics of individual layers of materials are important; however, there is limited work in the literature on the “single-layer” surface characteristics in SLM. Single-layer samples are designed on a smooth surface on top of semi-cylinders and fabricated using different levels of SLM processing parameters, i.e., laser power, scan speed, and hatch spacing, in different numbers of scan lines, i.e., 5, 10, 20, and 40. The fabricated sample surface data are acquired using non-contact optical interferometry. The computational approach utilizes the discrete element method (DEM) for the powder bed distribution, as well as computational fluid dynamics (CFD) for simulating the laser-powder interaction. One of the main contributions of this study is to investigate and provide predictive models for the correlations between the surface roughness of raster scans and SLM process parameters through experiments and computational work. In addition, a method for comparing the surface characteristics of the thermo-fluid simulation and experimental results is presented. A technique is employed to extract the surface roughness from the simulated results. The discussion on the experiment-simulation comparison challenges provides a deep insight into the criteria selection for experiment-simulation comparison in the SLM process. Moreover, an artificial neural network was trained using the back-propagation method and K-fold cross-validation, as well as regularization techniques to avoid overfitting. The laser power proved to be the most significant parameter, having a contribution of 48.3%, in determining the surface roughness with a unique trend affected by the spattering of powder particles. The simulation results showed noticeable deviations from the experiments in cases with extreme parameters, e.g., too high or too low values of power, speed, or hatch spacing. The machine learning algorithm achieves reasonable predictability of the single-layer surface roughness, showing a root mean square error and a coefficient of determination of 0.9589 µm and 98.78%, respectively, for a separate testing dataset.
