QUANTITATIVE ANALYSIS OF MACHINE LEARNING MODELS FOR DEFECT PREDICTION IN METAL ADDITIVE MANUFACTURING

Abstract

This study addresses the persistent challenge of predicting and controlling defects in metal additive manufacturing within an enterprise production environment, where porosity, lack of fusion, cracking, and geometric anomalies can increase scrap, rework, and qualification cost. The purpose was to quantify how technical and socio-technical factors jointly explain defect prediction effectiveness and deployment success, and to benchmark machine-learning models using risk-relevant performance measures. Using a quantitative, cross-sectional, case-based design, the sample combined (a) 162 usable survey responses from an enterprise AM facility workforce (quality engineering 32.1%, process engineering 26.5%, operations 23.5%, automation/data 17.9%; mean experience 6.8 years, SD 3.4) and (b) 312 inspected build or part cases verified via CT and metallography. Key variables included Data Quality, Feature Richness/Monitoring Capability, Process Window Stability, Technology Readiness, Interpretability/Trust, and Technology Acceptance Model constructs (Perceived Usefulness and Perceived Ease of Use), with the outcome measured as Defect Prediction Effectiveness and Deployment Success (Likert 1–5 composites). The analysis plan used descriptive statistics, reliability testing (Cronbach’s α = 0.82–0.90), Pearson correlations, and multiple regression, alongside build-wise model validation comparing Logistic Regression, SVM, Random Forest, and Gradient Boosting. Headline findings showed defects were dominated by porosity (46.8%) and lack of fusion (28.5%), with 12.9% classified as critical; drift-zone builds had higher defect prevalence (19.6%) than stable-zone builds (8.4%). Regression explained substantial outcome variance (R² = 0.61, Adj. R² = 0.59; F(7,154) = 34.78, p < .001), with significant effects for Data Quality (β = 0.19, p = .001), Interpretability/Trust (β = 0.21, p < .001), Perceived Usefulness (β = 0.29, p < .001), and other predictors. Gradient Boosting achieved the best predictive performance (Accuracy 0.89, F1 0.82, ROC-AUC 0.92) and the lowest false negative rate (10.8%), supporting risk-aware deployment. Implications are that enterprise defect prediction success depends not only on algorithms, but also on robust data governance, monitoring coverage, stable process windows, and explainable outputs that strengthen trust and workflow adoption.