Title:Toward data-driven research: preliminary study to predict surface roughness in material extrusion using previously published data with Machine Learning

Abstract:Material extrusion is one of the most commonly used approaches within the additive manufacturing processes available. Despite its popularity and related technical advancements, process reliability and quality assurance remain only partially solved. In particular, the surface roughness caused by this process is a key concern. To solve this constraint, experimental plans have been exploited to optimize surface roughness in recent years. However, the latter empirical trial and error process is extremely time- and resource-consuming. Thus, this study aims to avoid using large experimental programs to optimize surface roughness in material extrusion.
Methodology. This research provides an in-depth analysis of the effect of several printing parameters: layer height, printing temperature, printing speed and wall thickness. The proposed data-driven predictive modeling approach takes advantage of Machine Learning models to automatically predict surface roughness based on the data gathered from the literature and the experimental data generated for testing.
Findings. Using 10-fold cross-validation of data gathered from the literature, the proposed Machine Learning solution attains a 0.93 correlation with a mean absolute percentage error of 13 %. When testing with our own data, the correlation diminishes to 0.79 and the mean absolute percentage error reduces to 8 %. Thus, the solution for predicting surface roughness in extrusion-based printing offers competitive results regarding the variability of the analyzed factors.
Originality. As available manufacturing data continue to increase on a daily basis, the ability to learn from these large volumes of data is critical in future manufacturing and science. Specifically, the power of Machine Learning helps model surface roughness with limited experimental tests.