Accelerating discovery of glass materials in electronic devices through topology-guided machine learning

Abstract

The trial-and-error method, which is one of the most used methods in the glass industry, is extremely time-consuming and cost-intensive. In order to alleviate energy consumption and improve efficiency in glass development, a multi-tasking framework with topology-informed descriptors is proposed to obtain key properties, such as TEC, T_g, and T_m values, without conducting additional experiments. Herein, we disclosed that the casually generated findings on compositions are difficult to fabricate experimentally, and the predicted results largely deviate from the real values. To counter these issues, a confidence indicator was used to evaluate the composition's reliability, which embodied expert knowledge by classifying glass components into network formers, intermediates, and modifiers through the weighted Euclidean distance. Consequently, we were able to obtain real glass compositions with designed TEC values in the range of 3–12 ppm °C⁻¹ within a short cycle. Achieving a matched sealing with Kovar alloy (∼5.0 ppm °C⁻¹), five generated glasses exhibited TEC values closer to Kovar alloy and good wettability under 1000 °C. This work sheds light on developing efficient and cost-effective methods for the discovery of novel glasses in the field of electronic materials, such as GTMS, LTCC, and MLCC.