Accelerating the Exploration of Top Interface Passivators via Machine Learning for High-Performance Perovskite Solar...

Interface passivation at the perovskite/electron-transport-layer (ETL) is key to reducing defects in perovskite solar cells (PSCs), yet the broad chemical space of passivators hinders discovery. Here, we present a machine-learning (ML)-guided workflow accelerating the identification of effective ammonium-salt passivators for both n-i-p and p-i-n device architectures. A rigorously curated dataset of 296 literature records spanning 2017-2024, was stratified into 230 n-i-p and 66 p-i-n entries. Each passivator was represented via five molecular fingerprints and three descriptor sets, yielding eight distinct feature matrices per architecture. After training 72 regression models using tenfold cross-validated RandomizedSearchCV, 256&XGBoost performed best for n-i-p (test MAE = 0.0431, RMSE = 0.0647), while RDKit&SVR excelled for p-i-n (test MAE = 0.0393, RMSE = 0.0516.). SHAP revealed that passivator concentration, molecular weight within 120-280 g mol^-1, surface polarity fragments, and halide composition, particularly I/Br ratio, as the principal drivers of PCE enhancement. Guided by these insights, virtual screening of 162 salts at 9 concentrations identified top candidates. To validate our predictions experimentally, N,N,N-trimethyl-N-(3-hydroxypropyl) ammonium iodide was applied to a Cs_0.05FA_0.95PbI₃-based p-i-n device, yielding an improvement ratio of 1.152, closely matching the predicted value of 1.134 with a relative error of only 1.58%. Overall, this work establishes a transferable, SHAP-guided and architecture-aware machine-learning framework that efficiently screens top interface passivators and provides physically meaningful design rules across different device configuPassivator, machine learning, perovskite solar cellrations.