Machine learning for screening and predicting the best surface modifiers for a rational optimization of efficient perovskite solar cells

The key to keep the rising slope of perovskite solar cell performances is to reduce non-radiative losses by minimizing defect density. To this end, a large variety of strategies have been adopted spanning from the use of interfacial layers, surface modifiers, to interface engineering. Although winning concepts have been demonstrated, they result from a mere trial and error approach, which is time consuming and operator-dependent. To face this challenge, in this work, we propose the use of a machine learning approach for an educated and rational material screening with optimal characteristics in terms of surface passivation. In particular, we applied Shapley additive explanation to extract the specific chemical features of the passivator, which directly impact the device parameters, specifically the open circuit voltage (Voc). By monitoring the different material parameters as input, we were able to list the most promising passivators and directly test them in working solar cells. By comparing the device performances with the results of the modeling and with additional optical and morphological characterization, we retrieved the most significant material properties linked to the highest efficiency, which are (i) the presence of chlorine and its strong binding capacity to positively charged defects on perovskite surface, reducing the non-radiative recombination and (ii) an increased flexibility of the molecule, resulting in better coverage of the surface. Finally, we tested the predictive power of the ML algorithm proposing a new passivator, which, implemented in a working device, leads to the predicted high Voc confirming the results of the modeling.