Predicting Mechanical Properties of FDM-Produced Parts Using Machine Learning Approaches

Abstract

Additive manufacturing (AM), especially fused deposition modeling (FDM), has been widely used in industrial production processes in recent years. The mechanical properties of parts produced by FDM can be predicted through the correct selection of printing parameters. In this study, 25 machine learning (ML) algorithms were used to predict the mechanical properties (hardness, tensile strength, flexural strength, and surface roughness) of acrylonitrile butadiene styrene (ABS) samples fabricated by FDM. Experiments were conducted using three different layer thicknesses (100, 150, 200 μm), infill densities (50%, 75%, 100%), and nozzle temperatures (220°C, 230°C, 240°C). The effects of printing parameters on mechanical properties were investigated through analysis of variance (ANOVA). This analysis results indicated that infill density had the most significant effect on hardness (55.56%), tensile strength (80.02%), and flexural strength (77.13%). In addition, the layer thickness was identified as the most influential parameter on the surface roughness, with an effect of 70.89%. The prediction performance of the ML algorithms was evaluated based on the mean absolute error (MAE), root mean squared error, mean squared error, and R-squared (R2) values. The KSTAR algorithm best predicted both hardness and surface roughness, with MAE values of 0.006 and 0.009, respectively, and an R2 value of up to 0.99. For the prediction of tensile and flexural strength, the MLP algorithm was determined to be the most successful method, achieving high accuracy (R2 > 0.99) for both properties. In addition, comparison graphs between the predicted and actual results showed high overall accuracy, with a particularly strong agreement for hardness, tensile strength, and surface roughness. The study identified the algorithms with the best prediction performance and provided recommendations for predicting the 3D printing process based on these findings. © 2025 Elsevier B.V., All rights reserved.