Uncertainty Informed Screening for Safer Solvents used in the Synthesis of Perovskite Based Solar Cells via Machine Learning

The objective of this paper is to do a multi-output binary classification for endocrine disrupting (ED) nature of solvents that are frequently used in the synthesis of perovskites. Information on such solvents is not readily available in the form of datasets, rather it is embedded in the literature, which forms an ever-expanding corpus of scientific articles. Exploiting this corpus to extract relevant information on solvents is a mammoth undertaking and analyzing their ED nature is even more challenging. Except for a few solvents, little is known if they possess ED characteristics because of the resources required for in-vivo experiments. We address this challenge of expensive experiments by utilizing a deep-learning based model. In this work, using Natural Language Processing (NLP), we have identified 35 different organic solvents from a database of more than 30,000 paragraphs that are relevant to chemical synthesis of perovskites. Out of them, we have suggested 11 solvents as potential ED chemicals using a recently developed deep learning model. To further inform the quality of the classification, we perform an uncertainty quantification associated with the classification. This work serves as a guide in screening out the potential ED solvents, particularly when sufficient data is not available on them, thus paving the way for safer alternatives in perovskite synthesis.