Data‐Driven Molecular Encoding for Efficient Screening of Organic Additives in Perovskite Solar Cells

Abstract Machine learning (ML) has shown promise in screening organic molecular additives for planar perovskite photovoltaics, but is often hindered by predictive biases due to small datasets and reliance on predefined descriptors. Here, Co‐Pilot for Perovskite Additive Screener (Co‐PAS) is introduced, an ML‐driven framework designed to accelerate additive (or passivator) screening for perovskite solar cells (PSCs). Co‐PAS integrates the Molecular Scaffold Classifier (MSC) for scaffold‐based pre‐screening and utilizes Junction Tree Variational Autoencoder (JTVAE) to achieve data‐driven molecular structure representation, significantly enhancing the accuracy of power conversion efficiency (PCE) predictions. By applying Co‐PAS to screen 250 000 molecules randomly drawn from PubChem, candidates are prioritized based on predicted PCE values and key molecular properties, including donor number, dipole moment, and hydrogen bond acceptor count. This workflow helps narrow down to 76 promising candidates, including Boc‐L‐threonine N‐hydroxysuccinimide ester (BTN), a previously unexplored additive in PSCs. The solar cell with BTN achieves a device PCE of 25.20%. These results underscore the potential of Co‐PAS in advancing additive discovery for high‐performance PSCs.