Machine Learning-Assisted Materials Design and Discovery of Low-Melting-Point Inorganic Oxides for Low-Temperature Cofired Ceramic Applications

Abstract

The fabrication of low-temperature cofired ceramics (LTCCs) densified at a low sintering temperature (<900 °C) is energy-saving and environmentally friendly. However, finding novel LTCC materials by the trial-and-error method is time-consuming and costly. The LTCC materials often have low melting points, so it is feasible to discover high-performance LTCC materials out of the low-melting-point ceramics. A two-stage machine learning framework was adopted to establish the melting-point prediction model for inorganic oxides. Chemical compositions were used as features in stage 1 modeling; while in stage 2, more features were integrated according to domain knowledge to optimize the prediction model. Stage 2 model built by an artificial neural network algorithm shows the best performances with R2 = 0.7968 and root-mean-square error = 247.4 (K). Three features, including formation energy per atom (fepa), theoretical density (d), and number of atoms (na), were extracted as the decisive characteristics of inorganic oxides. The melting point demonstrates positive correlations with the absolute value of fepa and d. The na acts as a “recessive gene” because its contribution is indirect but necessary. The physical relationships between features and the melting point were also discussed. Furthermore, the LTCC inorganic oxides often have melting points lower than 1400 °C statistically. This criterion was verified by the reported LTCC/ultra-LTCC materials. The melting points of materials in the prediction set consisting of ∼3600 inorganic oxides were calculated by the ML model, and thus, the underlying LTCC materials could be screened out efficiently.