Machine-Learning-Based Analysis of Printing-Parameter Effects on Surface Roughness in FDM-Printed ULTEM 1010 Parts

Abstract

This study combines replicated experimentation and machine learning to characterize how user-controllable fused deposition modeling (FDM) parameters relate to local surface quality in complex ULTEM 1010 components. A surgical guide geometry was evaluated across 18 printing conditions incorporating the infill pattern, infill density, body thickness, raster angle, part orientation, and annealing. Three independently printed specimens per condition were measured at two locations, yielding 108 observations across five areal roughness metrics: Sa, Sz, Sq, Ssk, and Sku. An exploratory analysis of variance with false discovery rate correction identified orientation associations with Sa, Sq, Ssk, and Sku, and a body thickness association with Ssk in the cleaned measurements. Descriptive results associated the −XY orientation with a lower Sa and Sq at the measured regions, identifying a candidate placement for subsequent process trials. Evaluating multiple roughness metrics captured both the surface height magnitude and height distribution, providing a broader characterization than average roughness alone. Artificial neural network, random forest, and Gaussian process regression models were assessed using condition-grouped validation, which kept all replicates and paired measurement sites together and showed a limited generalization to unseen printing conditions. The study provides replicated evidence connecting industrially accessible printing settings with local areal surface characteristics. Its findings support prioritizing the orientation in process refinement, assessing the surface quality at functionally relevant locations, and validating predictive models on independent printing conditions before using them for parameter selection.