摘要
Additive engineering is an effective technique for alleviating intrinsic and interfacial defectsin perovskite light-emitting diodes (PeLEDs), yet discovering high-performance additivesremains challenging due to the vast chemical space and sparse experimental data. Here, weintroduce a generative deep learning framework to bridge this gap: leveraging a pre-trainedvariational autoencoder (VAE), we generate structurally similar additive candidates bysampling from the neighboring areas in the latent space of 142 known additives. This VAE-based approach, trained on large-scale molecule structure databases, mimics scientificintuitions of fine-tuning the existing molecular structure, and prioritizes chemical validityof the newly generated structure. Candidates are then screened via a multi-stage pipeline:an EQE prediction model, and descriptor-based evaluation (Gutmann’s donor number DN,Kamlet-Taft hydrogen-bond acceptor strength β) for defect passivation. We further validateour approach through experimental fabrication of FAPbI3 PeLEDs incorporating thegenerated additives. Among them, 1,3-dimethyltetrahydro-2(1H)-pyrimidinone (DMPU)leads to external quantum efficiency (EQE) of 16.0%. These results underscore theeffectiveness of using generative deep learning in discovery of molecular additive forperovskites.
危险品化学品经营许可证(带存储)