Data-Driven Design of Novel Halide Perovskite Alloys

TL;DR

This paper introduces a machine learning framework combined with high-throughput computations to design and predict properties of novel mixed cation halide perovskite alloys for optoelectronic applications.

Contribution

It develops a data-driven approach using neural networks and DFT calculations to efficiently screen and identify promising perovskite alloys with desirable properties.

Findings

Predicted 17,955 compounds' properties from a small DFT dataset.

Identified 574 promising perovskite alloys for photovoltaic applications.

Revealed compositional trends in mixed halide perovskites.

Abstract

The great tunability of the properties of halide perovskites presents new opportunities for optoelectronic applications as well as significant challenges associated with exploring combinatorial chemical spaces. In this work, we develop a framework powered by high-throughput computations and machine learning for the design and prediction of mixed cation halide perovskite alloys. In a chemical space of ABX$_{3}$ perovskites with a selected set of options for A, B, and X atoms, pseudo-cubic structures of compounds with B-site mixing are simulated using density functional theory (DFT) and several properties are computed, including stability, lattice constant, band gap, vacancy formation energy, refractive index, and optical absorption spectrum, using both semi-local and hybrid functionals. Neural networks (NN) are used to train predictive models for every property using tabulated elemental properties of A, B, and X site atoms as descriptors. Starting from the DFT dataset of 229 points, we use the trained NN models to predict the structural, energetic, electronic and optical properties of a complete dataset of 17,955 compounds, and perform high-throughput screening in terms of stability, band gap and defect tolerance, to obtain 574 promising compounds that are ranked as potential absorbers according to their photovoltaic figure of merit. Compositional trends in the screened set of attractive mixed cation halide perovskites are revealed and additional computations are performed on selected compounds. The data-driven design framework developed here is promising for designing novel mixed compositions and can be extended to a wider perovskite chemical space in terms of A, B, and X atoms, different kinds of mixing at the A, B, or X sites, non-cubic phases, and other properties of interest.