Machine learning and probabilistic analysis strategies for defect reduction in windshield manufacturing

• Machine learning ensemble achieves F1-score of 0.86 in defect classification. • Thinner glass (<5 mm) significantly increases risk of breakage in windshields. • Glass tone influences thermal stress patterns due to oxide composition. • Probabilistic analysis identifies critical thresholds for defect-prone variables. • Prescriptive strategies reduce defects through transport and handling protocols. Defects in automotive windshields (e.g., breakages, bubbles, cracks) compromise safety, increase production costs, and reduce product quality. This study integrates machine learning techniques and probabilistic modeling to optimize defect classification and derive prescriptive strategies for windshield manufacturing. A dataset of 1151 samples was analyzed through: (i) descriptive and correlational analysis of categorical (Thickness, Paint, Color, Tone, Sanding) and continuous (Weight, Area) variables; (ii) clustering and t-SNE for group identification; (iii) a stacking ensemble (MLP, Gradient Boosting, Random Forest, SVM, CatBoost) for classification; and (iv) probabilistic modeling with Shannon entropy for uncertainty assessment. Descriptive analysis revealed several important patterns. Area showed a bimodal distribution (IQR = 1.27–12.76 m², maximum = 39.89 m²), with Breaks more frequent in smaller surfaces and Cracked defects in larger ones ( H = 111.86, p < 0.001, η² = 0.024). Thickness <5 mm was linked to higher Breaks ( H = 19.16, padj = 1.8 × 10⁻⁴, η² = 0.022). Correlational analysis identified a moderate association between Tone and Color (Cramér’s V = 0.50), while Sanding was largely independent ( V ≤ 0.20). Kruskal–Wallis tests confirmed that Thickness ( H = 240.10, p < 0.001) and Tone ( H = 136.18, p < 0.001) were significantly associated with Weight and Area. Clustering ( k = 3, silhouette = 0.27) differentiated groups characterized by low weights/small areas with Breaks, larger areas with Cracked, and heavier compact units with Bubbles. The ensemble achieved a weighted F1-score of 0.83 on the hold-out test, with AUC = 0.96 (Breaks), 0.98 (Bubbles), 0.91 (Cracked). Its performance followed the same tendencies suggested by the exploratory phase, where nonlinear relationships between Thickness, Weight, and Area were already apparent. Building on this, probabilistic modeling refined the insights by detecting critical sensitivity thresholds: Breaks >60 % below 20 kg, with a high-risk zone at 11.3–17.2 kg (η² = 0.028), and Bubbles slightly increasing in heavy windshields >60–70 kg (≤15 %). Together, these results show a coherent narrative: exploratory analyses highlighted patterns, the model aligned with those trends, and probabilistic methods quantified the operational limits where defects become most likely. Based on this integrated analysis, three interventions were proposed: (i) structural reinforcement of thin glass (<5 mm), (ii) implementation of automated handling for large-area windshields, and (iii) ergonomic protocols to minimize operator-related defects. Overall, the complementarity between exploratory, modeling, and probabilistic approaches not only validated defect mechanisms but also provided prescriptive strategies for quality control. Nevertheless, limitations include the reliance on data from a single manufacturer, which may restrict generalizability, and the absence of real-time thermodynamic and pressure variables that could reveal additional hidden correlations. Future studies should expand datasets across multiple manufacturers and incorporate in-line physical measurements to enhance multivariable optimization and predictive control in automotive glass manufacturing.