Machine Learning Prediction of Heat Capacity for Solid Mixtures of Pseudo-Binary Oxides

Direct resistance-heated thermal energy storage in firebricks has been proposed as a low-cost energy storage alternative that could improve the economics of carbon-free electricity generation through both renewable and nuclear energy sources. The bricks used in these energy storage systems must be composed of oxide mixtures with specific material properties, amongst them a high heat capacity. While there are many methods available for the prediction of heat capacity in oxide mixtures, they all face different problems, with the Neumann-Kopp rule underperforming at higher temperatures and computational methods such as CALPHAD and Density Functional Theory being computationally expensive. Machine Learning algorithms have already shown promise in applications to predict the heat capacity of pure solid inorganics and could be used as a less computationally-demanding tool for the prediction of this property in pseudo-binary oxide mixtures, with the main impediment to its use being the unavailability of a heat capacity database for oxide mixtures. For this work, we use the CALPHAD software Thermo-Calc and its TCOX11 database of oxide Gibbs Free Energy functions obtained by approximation of empirical data in order to to generate a molar heat capacity dataset for pseudo-binary oxide mixtures at different temperatures and percentage compositions, and then use this dataset to train and test Machine Learning algorithms to predict molar heat capacity, using descriptors obtained from publically available DFT result repositories. By this process, we obtain algorithms capable of predicting heat capacity, some of them with an r² over 0.98, a MAE under 0.2 J/(molK) and a MAPE under 0.7%.