In Silico Materials Screening for Thermochemical Looping Applications Using Direct and Indirect Property Prediction Methods

Abstract

Thermochemical cycles involving the reversible reduction and oxidation of metal oxides are frequently investigated as a possibility for the conversion and storage of thermal energy from renewable sources. Heat provided by, for example, concentrated solar power (CSP) can be used to produce sustainable fuels using metal oxide-based thermochemical reaction cycles. Redox-active metal oxides have also been proposed for short-term thermochemical heat storage in CSP plants to ensure a continuous power generation despite the intermittency of solar energy. Among the materials previously reported for these applications, perovskite oxides are one of the most promising material classes for many types of thermochemical cycles. Nevertheless, further efficiency improvements are required for competitive solar-thermal fuel or electricity production at the heart of which is the need for novel high-performance oxide looping materials. While first-principles calculations using density functional theory (DFT) have become a ubiquitous tool in materials research for identifying novel materials with desired chemical or physical properties, the large compositional and structural variability of metal oxides makes a broad computational screening for thermochemical looping materials challenging. Furthermore, an accurate representation of the non-stoichiometric oxygen defect formation in perovskites requires large simulation cells not suitable in high-throughput calculations due to the prohibitive computational cost. To mitigate these issues, we have employed both direct and indirect property prediction methods in the search for novel oxide materials for thermochemical heat storage and air separation applications. Using existing and own DFT data, complex mechanical and thermodynamic oxide property estimates were inferred from simpler material features via direct prediction methods using existing regression and machine learning models. Furthermore, reported machine-learned interatomic force fields were used for indirect property predictions using them to calculate perovskite defect thermodynamics on a large scale and with high defect concentration resolution. The employed computational materials design approach allowed for a significantly accelerated candidate compound selection compared to a conventional DFT-only study. The identified materials can serve as rational starting points for further computational and experimental validation.