posted on 2024-07-10, 20:06authored byXunyong Yang, Yuqian Yang, Huimin Meng, Yu Li, Qin Hu
The rapid and accurate identification of potential high-efficiency
design strategies for perovskite solar cells (PSCs) is of paramount
importance in advancing their development and commercialization. However,
the application of machine learning (ML) algorithms in this field
is hindered by unstable PSC data sets (e.g., time-related noise and
data imbalance). Here, we introduce a ML framework specifically tailored
for temporal decoupling through feature engineering and uncertainty
modeling (noise processing techniques) to accurately predict the efficiency
of PSCs. In our framework, we utilize one-hot encoding and feature
fusion methods to extract features from a shared data set encompassing
perovskite material, processing, and PSC architecture. After temporal
decoupling, our ML model shows an outstanding precision of 96.88%
and a specificity of 99.3% for high-efficiency devices and is successfully
applied to predict PSC efficiencies in the period between 2021 and
2023. This temporal decoupling ML framework also reveals hidden relationships
between features and efficiency through cross-feature analysis. Our
work demonstrates the potential of ML for predicting performance and
elucidates the associated mechanisms, accelerating PSC commercialization
and reducing costs.