Implementing High‐Throughput Screening of Organic Solar Cells using Transfer Learning Based on Fine‐Tuning Neural Network Strategy

Abstract In organic solar cells (OSCs), traditional ensemble learning models have advanced the development of photovoltaic materials, reducing the reliance on labor‐intensive trial‐and‐error methods. However, these models suffer from insufficient generalization and poor transferability, leading to low accuracy in predicting power conversion efficiency (PCE) for new materials. In this work, a transferable neural network‐based framework is established to predict PCEs of binary OSCs. Specifically, 1431 sets of donor (excluding PM6):acceptor data are collected to train and validate four ensemble learning models and a transferable neural network model. These models achieved Pearson correlation coefficients ( r ) ranging from 0.75 to 0.84. Subsequently, a new dataset containing 113 sets of PM6:acceptor pairs is used to test their generalization abilities. The ensemble learning models exhibited significantly decreased r of 0.55–0.60, whereas the transferable neural network model maintained r above 0.80. Additionally, two electron acceptors differing only in their alkyl chain branching points are synthesized. The ensemble learning models predicted the similar PCEs for both acceptors. Conversely, the transferable neural network model predicted their significantly different PCEs, consistent with experimental results. This work demonstrates that the developed predictive framework offers substantial advantages in accurately predicting PCEs for new photovoltaic materials.