Identifying Performance Limiting Parameters in Perovskite Solar Cells Using Machine Learning

Abstract

Herein, it is shown that machine learning (ML) methods can be used to predict the parameter that limits the solar‐cell performance most significantly, solely based on the current density–voltage (J–V) curve under illumination. The data (11’150 J–V curves) to train the model is based on device simulation, where 20 different physical parameters related to charge transport and recombination are varied individually. This approach allows to cover a wide range of effects that could occur when varying fabrication conditions or during degradation of a device. Using ML, the simulated J–V curves are classified for the changed parameter with accuracies above 80%, where Random Forests perform best. It turns out that the key parameters, short‐circuit current density, open‐circuit voltage, maximum power conversion efficiency, and fill factor are sufficient for accurate predictions. To show the practical relevance, the ML algorithms are then applied to reported devices, and the results are discussed from a physics perspective. It is demonstrated that if some specified conditions are met, satisfying results can be reached. The proposed workflow can be used to better understand a device's behavior, e.g., during degradation, or as a guideline to improve its performance without costly and time‐consuming lab‐based trial‐and‐error methods.