Modified polyhedron model for predicting standard enthalpy of formation and entropy of mixed oxides

Thermodynamic modeling of oxidic systems is crucial in advancing various fields of science and technology. Polyhedron Model (PM) estimates the standard enthalpy of formation and entropy of mixed oxides via the linear summation of the thermodynamic properties of constituent polyhedra. Each polyhedron consists of a centered cation with neighboring oxygen anions; hence, the model accounts for the interaction between anions and cations. While second-order transitions have been considered in previous iterations of the model, the PM has certain shortcomings, including neglect of variations in polyhedron volume, polyhedron distortion, inter-polyhedron linkage, and second nearest-neighbor or higher-order interactions, which are not negligible. The present work introduces the Modified Polyhedron Model (MPM), which aims to incorporate these contributions through a neural network (NN) model to improve the accuracy of predictions for standard enthalpy of formation (${\Delta H}_{298K}^o$) and standard entropy ($S_{298K}^o$). This is possible by using the residuals from the PM as inputs to the NN model, whose outputs are the calculated thermodynamic properties of compounds. The dataset consists of 155 compounds in the Li-Na-K-Ca-Mg-Mn-Fe-Al-Ti-Si-O system, classified by 20 polyhedra. The MPM considerably reduces the error in predicting enthalpy of formation and entropy, improving the alignment with experimental values across most analyzed compounds in comparison with the PM. These results suggest that the MPM can significantly improve the predictability of thermodynamic properties for mixed oxides.