Interpretable glass type classification using genetic programming symbolic expressions based on chemical composition and refractive index

Classifying glass types from chemical composition and refractive index is essential for ensuring material quality, performance, and suitability across industrial applications, including construction, optics, and consumer products. In this study, we employ a Genetic Programming Symbolic Classifier (GPSC) to derive interpretable symbolic expressions (SEs) that distinguish between six glass types using the UCI glass identification dataset. Prior to model training, the data are preprocessed to address varying feature scales, limited sample size, and severe class imbalance. Multiple scaling/normalization schemes are combined with oversampling methods to generate balanced dataset variants, and a Random Hyperparameter Value Search (RHVS) is used to tune GPSC on each variant. To mitigate overfitting and enhance robustness, GPSC is trained with 5-fold cross-validation, and the resulting SEs are evaluated using standard metrics (ACC, AUC, Precision, Recall, F1-score). The best SEs per class are then fused into a Threshold-Based Voting Ensemble (TBVE) that aggregates their decisions via a class-specific threshold. On the original imbalanced benchmark setting, the resulting TBVEs achieve class-wise accuracies between 0.97 and 1.00 and AUC values above 0.96, clearly outperforming previously reported models on this dataset. When evaluated on an independent industrial dataset derived from real production compositions, all TBVEs maintain high performance, with class-wise F1-scores mostly above 0.93 and two classes achieving perfect discrimination. Detailed inspection of the learned SEs and a feature-importance analysis show that the dominant terms and variables align with established glass-chemistry knowledge, indicating that GPSC captures physically meaningful composition-property relationships rather than spurious correlations. Overall, the results demonstrate that symbolic classification provides a transparent yet flexible framework for glass-type identification with strong generalization to industrial conditions.