Downskin Surface Roughness Prediction with Machine Learning for As-Built CM247LC Fabricated Via Powder Bed Fusion with a Laser Beam

Abstract

Powder bed fusion with a laser beam (PBF-LB) is a widely used metal additive manufacturing method for fabricating complex three-dimensional components with a variety of metallic powders. However, metal parts fabricated by PBF-LB often present surface quality problems because of the layer-wise building process and the occurrence of partially unmelted powder particles. To reduce the surface roughness, surface post-processing is required, which incurs additional time and cost. In particular, the downskin surface generally has the worst surface roughness among the fabricated components. The rough surface reduces the lifetime and quality of the holed part owing to cracks, corrosion, and wear. In this study, for fast and efficient improvement of the downskin surface roughness of CM247LC fabricated by PBF-LB, machine learning algorithms, namely support vector regression (SVR), random forest (RF), and multilayer perceptron (MLP), were introduced to predict downskin surface roughness in the process parameter selection step. Three PBF-LB process parameters (laser power, scanning speed, and hatching distance) and the overhang angle were selected as the input variables for the machine learning models for predicting downskin surface roughness. Test samples were prepared and used for training and evaluation of the proposed machine learning algorithms, with RF showing the most promising results. Early results were confirmed when model predictions were compared to the actual measured roughness of a fabricated vane part, with average deviations of 13.7%, 4.3%, and 22.5% observed for SVR, RF, and MLP, respectively. The results showed that the proposed machine learning models could accurately predict the downskin surface roughness in the process parameter selection step without the use of any sensor, with RF showing the highest prediction accuracy.