Aplicação de métodos de seleção de características em dados de vidros

Abstract

Oxide glasses are non-crystalline compounds with disordered structures, which give rise to unique properties applicable across various fields. Moreover, there is significant potential for discovering new glass-forming structures and several studies have applied ways of predicting the physical properties of glasses (ALCOBAÇA, 2020; LIU; SU, 2024). In this context, machine learning models have emerged as a promising approach for identifying novel glasses with desirable properties and compositions for diverse applications. However, due to the sparse nature of glass composition data, it is crucial to first understand how chemical elements influence specific physical properties - such as the glass transition temperature (Tg), a key indicator of the transformation from solid to viscous state - before developing highly accurate predictive models. This study aims to comparatively analyze three feature selection algorithms - Random Forest (RF), Recursive Feature Elimination (RFE) applied to Support Vector Regressor (SVR), and Linear Regression (LR) - for predicting Tg using the SciGlass oxide glass database. After preprocessing and dimensionality reduction to 37 chemical elements using the Maximum Information Coefficient (MIC) and Pearson Correlation Coefficient (PCC), the models were trained, tested, and quantitatively validated using Shapley Additive exPlanations (SHAP). The most relevant elements identified by the algorithms were interpreted considering glass science. SHAP enhanced the interpretability of RF predictions and revealed that elements identified by all three methods - Bismuth, Lead, Tellurium and Vanadium - tend to reduce Tg by compacting the network, while elements selected by RF and RFE - Aluminum, Calcium, Lanthanum, and Titanium - increase Tg by enhancing glass rigidity. These findings are aligned with the experimental CaO–Al₂O₃–SiO₂ system used in oxide glass production. This research contributes to advancing chemical element selection strategies and supports the development of more accurate predictive models in the future.