Computation of photovoltaic and stability properties of hybrid organic–inorganic perovskites via convolutional neural networks

Abstract

Hybrid organic–inorganic perovskites (HOIPs) have gained considerable interest due to their potential applications as photovoltaic materials. Nevertheless, several issues have to be solved on this matter, such as the proper tuning of band gaps and those concerning stability, before these systems can realise their full potential. Here, we used deep learning techniques, more specifically crystal graph neural networks (Xie & Grossman, Phys. Rev. Let., 2018, 120), abbreviated as CGNN, to explore the chemical space of HOIPs and to address the above mentioned difficulties. We trained this CGNN with a data set comprised of 1346 density functional theory calculations and used it to compute band gaps, refractive indexes, atomisation energies, volumes of unit cells and volumetric densities of 3840 HOIPs. Our screening method permits a rapid selection of perovskites with suitable optoelectronic properties and only 7 have an adequate band gap to be used in photovoltaic technologies. The composition, ABX$_3$, of such perovskites is mainly of small molecular cations such as A = $\mathrm {[NH_4]^+}$, $\mathrm {[NH_2NH_3]^+}$ together with $\mathrm {[OHNH_3]^+}$, B = $\mathrm {In^2+}$, $\mathrm {Zr^2+}$ along with $\mathrm {Sn^2+}$, and X = I$^-$. The consideration of further systems indicates that the occurrence of phosphorus and sulphur in the molecular cation diminishes strongly the band gap of the perovskite. We also considered the stability of the systems with optimal band gaps with respect to their degradation in simple organic and inorganic salts. Overall, our investigation shows how deep learning techniques can be exploited to achieve a rapid screening of potential photovoltaic materials in terms of their electronic properties and stability.