Self-Supervised Learning for Glass Composition Screening

Glass composition screening seeks to identify the optimal combination of components that satisfy specific performance criteria, thereby facilitating the development of new glass materials for a wide range of applications. However, the inherent complexity of multicomponent glass systems presents significant challenges in effectively correlating composition with property and achieving efficient screening processes. Current machine learning approaches for composition screening predominantly rely on supervised learning paradigms, which focus on precisely fitting sample labels. These methods not only require large amounts of high-quality data but are also susceptible to overfitting noisy samples and frequently occurring head samples, thereby limiting their generalization capabilities. In this study, we introduce a novel self-supervised learning framework specifically designed to screen glass compositions within predefined glass transition temperature (Tg) ranges. We reformulate the compositional screening task as a classification problem, aiming to predict whether the Tg of a given composition falls within the specified range. To enhance the training dataset and improve the model's resilience to noise, we propose an innovative data augmentation strategy grounded in asymptotic theory. Additionally, we present DeepGlassNet, a specialized network architecture engineered to capture and analyze the intricate interactions between different glass components. Our results demonstrate that DeepGlassNet achieves significantly higher screening accuracy compared to traditional methods and is readily adaptable to other composition screening tasks.