Accelerating Identification of Relevant Polarons in Batteries

Abstract

Many of the materials under consideration as battery electrodes are known to exhibit polarons─charge carriers localized through polarizing their environments, which are unfortunately quite costly to tackle by theory. Next to costly ab initio methods and geometry optimizations, the main problem is the sheer number of possible configurations polarons can adopt in a material. This problem is exacerbated by symmetry-breaking defects such as oxygen vacancies ubiquitous in oxide materials. Fortunately, a majority of polaron configurations tend to be too high in energy to ever be adopted at normal operating temperatures. In this work, we therefore present a data-driven approach based on gradient-boosted decision trees as a descriptor for relative polaron energies. We demonstrate the efficacy of this approach to weed out those configurations that are energetically unlikely in three simulation cells of defected spinel Lithium₄Titanium₅Oxygen₁₂ (LTO). The first two cells differ by their 16d-site occupational disorder and the position of the oxygen vacancy, which in both cases leads to the formation of two polarons in the cells. In contrast, the third simulation cell contains an extra lithium ion, which leads to three polarons. Our approach allows us to reduce the number of necessary first-principles calculations by roughly two-thirds. It can easily be adapted to other materials and simulation cells, greatly speeding up their theoretical treatment.