Surface Roughness Prediction for Additively Manufactured Ti-6Al-4V Components Based on Supervised Learning Models

Abstract To avoid redundant trial and error experiments in hope of achieving acceptable surface roughness, reliable predictive models must be engineered to anticipate surface characteristics based on process parameter inputs. In the present study, two rigorously tested supervised machine-learning based models are proposed to predict the arithmetic average (Ra) of profile deviation for the surface of Ti-6Al-4V alloy manufactured via selective laser melting (SLM). Firstly, a Gaussian Process Regression (GPR) model with Rational Quadratic kernel function is constructed after ten-fold cross validation of the training data. Secondly, using the same training data and the same ten-fold cross validation, a feed-forward narrow neural network (NN) is employed. Primary input parameters of SLM process, namely laser power, scanning speed, hatch spacing, layer thickness and volumetric energy density are mined from literature investigating as-built surface characteristics of SLMed Ti-6Al-4V alloy. To further test the developed machine-learning models, ten 8 × 8 × 8 mm Ti-6Al-4V cubes are manufactured and a comparative between two non-parametric (GPR and NN) models is performed by predicting the surface roughness (Ra) of the ten samples. It is discovered that the NN model underperforms with a root mean squared error (RMSE) of 2.76 μm, as opposed to its counterpart GPR model, exhibiting RMSE of 0.82 μm. Additionally, analyses of the surface characteristics of the fabricated samples, the surface profiles, and impact of such a predictive model are provided.