Addressing accuracy by prescribing precision: Bayesian error estimation of point defect energetics

With density functional theory (DFT), it is possible to calculate the formation energy of charged point defects and in turn to predict a range of experimentally relevant quantities, such as defect concentrations, charge transition levels, or recombination rates. While prior efforts have led to marked improvements in the accuracy of such calculations, comparatively modest effort has been directed at quantifying their uncertainties. However, in the broader DFT research space, the development of Bayesian Error Estimation Functionals (BEEF) has enabled uncertainty quantification (UQ) for other properties. In this paper, we investigate the utility of BEEF as a tool for UQ of defect formation energies. We build a pipeline for propagating BEEF energies through a formation-energy calculation and test it on intrinsic defects in several materials systems spanning a variety of chemistries, bandgaps, and crystal structures, comparing to prior published results where available. We also assess the impact of aligning to a deep-level transition rather than to the VBM (valence band maximum). We observe negligible dependence of the estimated uncertainty upon a supercell size, though the relationship may be obfuscated by the fact that finite-size corrections cannot be computed separately for each member of the BEEF ensemble. Additionally, we find an increase in estimated uncertainty with respect to the absolute charge of a defect and the relaxation around the defect site without deep-level alignment, but this trend is absent when the alignment is applied. While further investigation is warranted, our results suggest that BEEF could be a useful method for UQ in defect calculations.