Probabilistic deep learning approach for targeted hybrid organic-inorganic perovskites

We develop a probabilistic machine learning model and use it to screen for new hybrid organic-inorganic perovskites (HOIPs) with targeted electronic band gap. The dataset used for this work is highly diverse, containing multiple atomic structures for each of 192 chemically distinct HOIP formulas. Therefore, any property prediction on a given formula must be associated with an irreducible ``uncertainty'' that comes from its unknown atomic details. We show that the probabilistic deep learning approach supported by TensorFlow Probability library, is robust, versatile, and can properly capture the aforementioned uncertainty. Dozens of new HOIP formulas with band gap falling between 1.25 eV and 1.50 eV were identified and validated for against suitable first-principles computations. We argue that data uncertainty, which is inevitable in materials informatics, should be handled properly, and towards this goal, probabilistic deep learning is a suitable and reliable approach.