Discovering Novel Halide Perovskites using Multi-Fidelity Machine Learning and Graph Neural Networks

The ABX₃ perovskite crystal structure is ubiquitous and the subject of extensive study owing to the sheer tunability of electronic and optical properties that can be achieved. Halide perovskites, in particular, are materials of great interest for solar absorption and many related optoelectronic applications such as LEDs, lasers, and UV or IR sensors. The discovery of novel perovskite compositions, including complex alloys with attractive properties, is hindered by the combinatorial nature of the chemical space and a general lack of quantification of systematic inaccuracies in simulations such as from first principles-based density functional theory (DFT). In this work, we generated large datasets of computed stability, electronic band gaps, theoretical photovoltaic efficiency derived from optical absorption spectra, and defect formation energies, of halide perovskite alloys from various DFT semi-local and hybrid functionals. This data is combined with smaller quantities of corresponding experimental measurements collected from the literature, and used for training multi-fidelity machine learning (ML) models to make property predictions at experimental accuracy. Such predictions, using state-of-the-art nonlinear regression techniques including random forests and Gaussian processes, are sequentially improved and coupled with a recommendation engine for new computations and experiments to gradually achieve new stable compositions with targeted band gap and absorption. Initial success if obtained using as inputs compositional information, known elemental properties of A, B, and X species, and one-hot encoding of perovskite phase and data fidelity. We extend these models to a much larger dataset of &gt; 10,000 perovskite structures, wherein entire crystal structures are used as input via a variety of graph neural network-based approaches. This enables inclusion of lattice strain, octahedral distortions and rotations, and different kinds of alloy ordering as implicit inputs to the ML framework, ultimately resulting in general models applicable to any atom-composition-structure combination within the selected halide perovskite chemical space. Best DFT-ML surrogate models are coupled with optimization schemes using methods such as genetic algorithm, used to drive collaborative experiments and further DFT computations, and made available to the community via user-friendly tools on the NSF-funded online repository, nanoHUB.<br/> <br/><b>References</b><br/>1. A. Mannodi-Kanakkithodi et al., "Data-Driven Design of Novel Halide Perovskite Alloys", <i>Energy Environ. Sci.</i> 15, 1930–1949 (2022).<br/>2. J. Yang et al., “High-throughput computations and machine learning for halide perovskite discovery”, <i>MRS Bulletin.</i> (2022). https://doi.org/10.1557/s43577-022-00414-2<br/>3. T. Xie et al., “Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties”, <i>Phys. Rev. Lett.</i> 120, 145301, (2018).<br/>4. C. Chen et al., “Graph Networks as a Universal ML Framework for Molecules and Crystals”, <i>Chem. Mater.</i> 31 (9), 3564-3572 (2019).<br/>5. K. Choudhary et al., “Atomistic Line Graph Neural Network for improved materials property predictions”, <i>npj Comput Mater.</i> 7, 185 (2021).