Machine learning optimization for a novel process of acid-base chemical densification of BaZrO3 ceramics.

Abstract

Low-temperature ceramic fabrication techniques, which can reduce CO2 emissions and energy consumption, are important. BaZrO3 ceramics, which are conventionally sintered at over 1600 ºC, were prepared at close to room temperature by a novel process called acid-base chemical densification (ABCD). BaZrO3 ceramics could be fabricated by reacting a compact of ZrO2 hydrous gel with a solution of Ba(OH)2•8H2O at less than 100 ºC. The densities of BaZrO3 ceramics depended on the various ABCD process parameters. To obtain a higher density of BaZrO3 ceramics, all the parameters were optimized simultaneously using machine learning. Three algorithms, Linear Regressor (LR), Random Forest (RF) regressor, and Light Gradient Boost Machine (LGBM), were compared for the prediction of BaZrO3 density. The RF regressor predicted the relative densities of BaZrO3 ceramics using the process parameters with high accuracy. From the results of the important features obtained from the RF model, the water content in the zirconia hydrous gel used as a precursor is important for the densification of BaZrO3 in this process, although the reaction time and precursor density affected the densities in the ABCD process. In contrast, varying the reaction temperature in the chemical reaction between ZrO2 hydrous gel and Ba(OH)2·8H2O did not effectively improve the density in this study. By analyzing the top six important features in the RF model, we attempted to improve the accuracy of density prediction using the LR to predict the densities from the extrapolation data. Using an improved LR model, it was found that reducing the water content in the precursors is important for obtaining dense BaZrO3 ceramics by the ABCD process. The water in the gel disturbs the densification because the volume of water molecules remains as pores in the pellets.