Orthorhombic Structures as High-Throughput Models for Inorganic Halide Perovskites

Inorganic halide Perovskite’s (HP) low-cost solution processing, defect tolerance, long carrier lifetimes, and band gap tunability via compositional engineering have propelled interest in their use as light absorbers in photovoltaic (PV) devices. The perovskite PV literature is primarily populated by lead (Pb) containing perovskites with compositions such as CsPbI₃ having taken the PV community by storm due to their exceptional PV performance. However, recently there has been a push towards discovering new HP compositions that may offer superior performance while moving away from the health and safety concerns commonly associated with Pb.<br/>Engineering new HP compositions by substituting different elements for Cs on the A site, Pb on the B site, and up to 4 halides on the X site yields approximately 10⁴ theoretical ternary (ABX3) compositions. Moreover, the HP materials space explodes to upwards of 10⁸ possible compositions when considering double and alloyed perovskite compositions. These spaces are too large to characterize with experimental methods alone, motivating the use of density functional theory (DFT) and machine learning (ML) for materials screening.<br/>The field has used different crystal structures, such as the primitive cubic structure and polymorphous networks (PN), to compute perovskite properties using DFT. The former is a 5-atom unit cell that is computationally inexpensive to calculate but often severely underestimates experimental band gaps; conversely PNs use large unit cells (160 atoms) that yield good band gap agreement with experiment but are computationally expensive to calculate. This makes neither model suitable for HT screening. Hence, we propose the orthorhombic perovskite structure as a surrogate model that strikes a balance between accuracy and cost. The surrogate model is a medium sized cell of 20 atoms that we initialized with stabilizing orthorhombic octahedral tilting, rescaled the lattice, and displaced the atomic sites to help maintain a realistic representation of the perovskite structure prior to DFT optimization. Our surrogate model provides a ~40x speed up in calculations compared to PNs, whilst being able to replicate PN band gaps with a MAD of only 0.09 eV. The orthorhombic surrogate model is useful for constructing large datasets of theoretical perovskite properties to enable HT perovskite materials screening and provide large sets of high quality DFT data for data mining and machine learning applications.