Advancing the prediction of crystalline phases in glass-ceramics via machine learning

Abstract

Lithium aluminosilicate (LAS) glass-ceramics are widely utilized in diverse application, owing to their outstanding properties such as high transparency, high fracture toughness and ultra-low thermal expansion. The development of LAS glass-ceramics has traditionally relied on the phase diagram to identify primary crystalline phases. However, this approach is limited by constrained composition ranges and unknown heating treatment parameters, hindering the efficient development of high-performance glass-ceramics. In this study, we establish a comprehensive small-scale database of LAS glass-ceramics, comprising 751 samples characterized by 27 compositions, nucleation temperature, nucleation time, crystallization temperature, crystallization time and 13 crystalline phases. We employ five algorithms, i.e. Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Classification and Regression Trees (CART), K-Nearest Neighbors (K-NN) and Multi-Layer Perceptron (MLP) Classifier to predict the potential crystalline phases. Our results demonstrate that RF achieves the best overall performance, with the highest accuracy of 0.8609, the lowest hamming loss of 0.0142, and the highest micro F1 score of 0.9234. This work advances the understanding and prediction of crystalline phases in LAS glass-ceramics, providing valuable insights for the development and optimization of these materials.