QSPR‐Driven Data Augmentation for Machine Learning Models: Prediction of Glass Chemical Durability

ABSTRACT

The chemical durability of glass is a critical property governing its performance in advanced applications, including optical communication, biomedical devices, and nuclear waste immobilization. However, the intricate composition–structure–property relationships in amorphous materials have traditionally necessitated reliance on empirical design and inefficient trial‐and‐error methods. To transcend these limitations, computational and data‐driven approaches have become indispensable tools for accelerating the discovery and optimization of glass compositions. In this work, a QSPR‐driven data augmentation strategy has been developed to construct accurate machine learning models for chemical durability prediction with limited experimental data. The established QSPR‐χ methodology was adopted to construct the foundational structure–property model, and it was subsequently employed to enhance the property datasets. In addition, systematic development and evaluation were performed on a suite of algorithms and methodologies. The results show that the data partitioning strategy have negligible impact on the generalization performance of the models. Furthermore, Gaussian Process Regression (GPR) demonstrates superior predictive capability and robustness compared with other algorithms. Moreover, the interpretability analysis of the optimal models elucidates that the roles of specific oxide components, that is, Al ₂ O ₃ , B ₂ O ₃ , and CaO, play pivotal factors in controlling the chemical durability of the studied glass systems. This study establishes a QSPR‐driven data augmentation for constructing machine learning models with limited data points, which provides a powerful tool for the rapid prediction of glass chemical durability.