Data-driven design of halide perovskites using high-throughput computations and machine learning

Halide perovskites are highly desirable for optoelectronic applications due to their extraordinary tunability, especially via composition engineering and octahedral rotation and distortion. However, significant challenges arise when exploring combinatorial possibilities for atoms, compositions and structures, which results in a computationally and experimentally intractable problem. In this work, we address this issue using a combination of high-throughput density functional theory (DFT) and machine learning (ML). We generate a DFT dataset of ~ 550 perovskite alloys based on a selected set of A, B, and X atoms and arbitrary mixing at each site, calculating several critical properties using standard GGA-PBE and hybrid HSE06 functionals, including band gap, decomposition energy, photovoltaic absorption, and vacancy formation energy. Predictive models are trained from this dataset using descriptors that encode compositional, elemental, and octahedral information, and applying state-of-the-art ML regression techniques such as random forests and neural networks. These DFT-surrogate models are used for screening of thousands of promising new candidates and combined with a genetic algorithm framework for inverse design of new compositions with multiple targeted properties, which are validated with additional computations and recommended for experimental synthesis and characterization. The large datasets and AI-based prediction and optimization schemes resulting from this work are promising for the accelerated design of novel halide perovskites for solar cells, photodiodes, electronics, infrared sensors, and other related applications.