Assessing Perovskites' Stability Through Machine Learning Algorithms

Halide perovskites have high-performing yet variable optical properties, which is currently impeding their adoption in optoelectronic devices. The exact degradation mechanisms are not well understood due to the immense composition parameter space involved¹², and their dynamic, often confounding response to environmental stressors (such as humidity, temperature, oxygen, light, and bias).³ To quantify the degradation of several different perovskite compositions, we use a custom-built, high throughput, <i>in situ</i> photoluminescence (PL) characterization to collect sufficient data to train a variety of machine learning (ML) models. We compare forecasts of PL of several Cs<i>_y</i>FA<i>_1-y</i>Pb(Br<i>_x</i>I<i>_1-x</i>)<i>₃</i>perovskites in response to relative humidity and temperature cycling with linear regression (LR), echo state network (ESN), and seasonal autoregressive integrated moving average with exogenous regressors (SARIMAX)⁴, and extreme gradient boosting (XGBoost)⁵ algorithms. Overall, we find that these models enable predictions of PL figures of merit, such as peak location, area, intensity, and full-width half max (FWHM), with up to 98% accuracy when trained on 100 hours of data. Furthermore, we create a generalized model that can predict the PL behavior of ten compositions unseen during model training. The relative feature importance and correlations between environmental inputs and optical performance show that composition dominates sample degradation. The stable compositions, which in this family contain small percentages of cesium and bromine, respond more cyclically to environmental stressors, and they recover their optical properties after rest. The development of automized workflows for high- throughput data collection and ML analysis greatly accelerates energy materials discovery, as this model can be extended to encompass up to hundreds of perovskite compositions.⁶<br/><br/><br/><br/><br/>(1) Srivastava, M.; Howard, J. M.; Gong, T.; Rebello Sousa Dias, M.; Leite, M. S. Machine Learning Roadmap for Perovskite Photovoltaics. <i>The Journal of Physical Chemistry Letters</i> <b>2021</b>, <i>12</i> (32), 7866–7877. https://doi.org/10.1021/acs.jpclett.1c01961.<br/>(2) Howard, J. M.; Wang, Q.; Srivastava, M.; Gong, T.; Lee, E.; Abate, A.; Leite, M. S. Quantitative Predictions of Moisture-Driven Photoemission Dynamics in Metal Halide Perovskites via Machine Learning. <i>J. Phys. Chem. Lett.</i> <b>2022</b>,<i>13</i> (9), 2254–2263. https://doi.org/10.1021/acs.jpclett.2c00131.<br/>(3) Boyd, C. C.; Cheacharoen, R.; Leijtens, T.; McGehee, M. D. Understanding Degradation Mechanisms and Improving Stability of Perovskite Photovoltaics. <i>Chem. Rev.</i> <b>2018</b>, <i>119</i>. https://doi.org/10.1021/acs.chemrev.8b00336.<br/>(4) Srivastava, M.; Hering, A. R.; An, Y.; Correa-Baena, J.-P.; Leite, M. S. Machine Learning Enables Prediction of Halide Perovskites’ Optical Behavior with &gt;90% Accuracy. <i>ACS Energy Lett.</i> <b>2023</b>, <i>8</i> (4), 1716–1722. https://doi.org/10.1021/acsenergylett.2c02555.<br/>(5) Hering, Abigail R; Dubey, Mansha; Imeni, Seyede Elahe Hosseini; Srivastava, Meghna; An, Yu; Correa-Baena, Juan-Pablo; Homayoun, Houman; Leite, Marina S. Machine Learning Reveals Composition Dependent Properties in Halide Perovskite. <i>manuscript in progress</i> <b>2024</b>.<br/>(6) Hering, A. R.; Dubey, M.; Leite, M. S. Emerging Opportunities for Hybrid Perovskite Solar Cells Using Machine Learning. <i>APL Energy</i> <b>2023</b>, <i>1</i> (2), 020901. https://doi.org/10.1063/5.0146828.