A Priori Data Collection for Thermodynamic Modeling of Off-Stoichiometric Metal Oxides Via Bayesian Methods
Steven Wilson, Arizona State University
The thermodynamic properties of metal oxides as a function of off-stoichiometry are crucial materials design attributes in metal oxide-based oxygen-exchange chemical processes. It has been shown that the compound energy formalism (CEF) is a powerful framework to describe these thermodynamic properties. Recently, we developed method for fitting the CEF model that overcomes the current fitting challenges associated with non-unique fits in part, due to the thousands of possible liner combinations of the parameters involved resulting in dynamic behavior between enthalpy and entropy trends. The key innovations are: 1) the combination of density functional calculations with experimental data that delineates the enthalpic/entropic contributions to the Gibbs free energy; 2) a systematic determination of the important CEF model terms, removing thermodynamic predetermining human intervention; 3) a self-consistent solution of the starting oxygen off-stoichiometry (δ0) of thermogravimetric measurements. With the advent of this state-of-the-art algorithm our method enables the reliable extraction of off-stoichiometric metal oxide thermodynamic properties and facilitates rapid materials compositional screening, and reliable process design of systems dependent on off-stoichiometric redox-active metal oxides. We build on this concept by examining the use of a Bayesian approach to selecting data points which should be gathered next. Data points to examine include measuring the non-stoichiometric at a particular temperature and pressure in an experimental dataset and new theoretically calculated points, i.e. the energy of a particular system composition and off-stoichiometry. The combination of these methods provides more reliable thermodynamic characterization and identifies, a priori, critical points to examine in the temperature, O2 partial pressure, compositional landscape, thus minimizing the number of points which must be examined. Specifically, we show that a model identical to a ground truth model can be realized with half the data from careful selection of data points through the Bayesian approach.
Abstract Author(s): Steven A. Wilson