Graph Neural Network-Assisted Machine Learning for High-Throughput Buried Interfacial Materials Screening in...

Numerous interfacial materials have been explored for efficient perovskite solar cells (PSCs). The identification of optimal candidates from the vast chemical space remains a difficult and costly task. Artificial intelligence (AI)-assisted materials screening has emerged as a convenient route to benefit and promote the pace of PSC research. Here, we propose a two-step cascade screening strategy that integrates a contrastive learning-based graph neural network (CLGNN) with a gradient boosting decision tree regressor (GBDTR) to balance high-throughput and predictive accuracy. Among them, CLGNN is a graph contrastive learning model trained on large-scale unlabeled molecular graph topologies, whereas GBDTR is a supervised regression model trained on a manually curated and labeled database. Using CLGNN, we screened one million molecules on a graphics processing unit (GPU) and shortlisted 128 candidates within 10 min. This step substantially reduced the number of molecules requiring subsequent density functional theory (DFT) calculations and experimental validation. We then developed and evaluated multiple regression models based on multidimensional physicochemical descriptors to accurately predict device performance. Two effective buried-interface modifiers, 4,4'-iminodibenzoic acid (4,4'-IDA) and 4,4'-biphenyldicarboxylic acid (4,4'-BA), were selected, which delivered a high-power conversion efficiency of 26.10% with an open-circuit voltage (V_OC) of 1.184 V and a fill factor (FF) of 86.05% using antisolvent-free processing. This work exemplifies an efficient route for scalable and rapid exploration of new materials for PSCs.