Extracting processing and testing parameters from materials science literature for improved property prediction of glasses

Abstract

Recent literature has seen a surge in artificial intelligence (AI) and machine learning (ML) for computational design and discovery of novel materials. Most of these studies focus on inverse design or property prediction through composition–structure–property relationships. However, a crucial aspect that has received less attention is the materials' processing route and testing conditions. Here, we present a framework combining text-mining with ML, enabling extraction of processing and testing conditions in addition to the material composition for improved property prediction. We demonstrate this framework by predicting the hardness of inorganic glasses with composition, annealing temperature, and loading conditions as input. Specifically, we analyze the text associated with research articles to obtain the process parameters of oxide glasses with various input components. We show that process parameters and composition together as input yield superior predictions of hardness. Interpreting the predictions using a game-theoretic approach, namely SHAP, we show that the processing and testing parameters indeed play a significant role in controlling the hardness of glasses. Finally, we show that the proposed approach can outperform heuristic approaches for glass synthesis, allowing scientists to choose the most appropriate synthesis route for developing glasses with targeted properties.