Developing an AI-Based Digital Twin for Predictive Quality Control in 3D Printing Processes

Abstract

Additive manufacturing can produce complex parts, yet print quality often changes between runs. Small changes in temperature, speed, material, or layer settings may cause rough surfaces and weak parts. This paper develops an AI-based digital twin for predictive quality control in 3D printing. The twin joins machine data, process settings, quality models, and operator decisions within one synchronized system. It predicts surface roughness, tensile strength, and elongation before final inspection. A public material-extrusion dataset provides the practical test. The dataset contains 70 printing runs from an Ultimaker S5 printer. Nine process settings and three measured quality outcomes are available. One invalid record was removed before model training. Ridge regression, random forest, gradient boosting, and extra trees were evaluated. Repeated five-fold cross-validation reduced dependence on one favorable data split. Gradient boosting achieved the lowest average normalized error across the three outcomes. Its out-of-fold R² reached 0.90 for roughness, 0.53 for tensile strength, and 0.54 for elongation. The mean absolute errors were 21.49 µm, 4.50 MPa, and 0.39%, respectively. A replay test converted these predictions into simple quality alerts. The alert logic reached 79.1% balanced accuracy under illustrative acceptance limits. Results show useful predictive value, but they do not prove industrial readiness. Larger datasets, sensor streams, uncertainty limits, and printer-specific validation remain necessary. The proposed structure offers a practical path from offline machine learning toward safe, traceable, closed-loop quality control.