Development of organic aggregation-induced emission fluorescent materials based on machine learning models and experimental validation

• Published in: Journal of molecular structure 2024-12, Vol.1317, p.139126, Article 139126
• Main Authors: Zhao, Yihuan, Chen, Kuan, Yu, Binhai, Wan, Qianwen, Wang, You, Tang, Fushan, Li, Xinmin
• Format: Article
• Language: English
• Subjects: AIE materials; Machine learning; Organic fluorescent materials
• ISSN: 0022-2860
• DOI: 10.1016/j.molstruc.2024.139126

•A dataset comprising 3074 AIE/ACQ molecules was constructed through an extensive literature search.•The LightGBM algorithm was utilized to develop a machine learning (ML) model for predicting the AIE/ACQ properties of organic fluorescent materials.•The extrapolation ability of the ML prediction model was further validated through out-of-sample and experimental validation. Organic fluorescent materials have attracted significant attention due to their unique chemical and optoelectronic properties. Unfortunately, traditional organic fluorescent materials are limited by aggregation-caused quenching (ACQ), which restricts their application potential in various fields. The discovery of aggregation-induced emission (AIE) materials has revolutionized the field by providing a solution to this issue. Despite considerable efforts to design and synthesize novel AIE materials, the development process still relies on tedious trial-and-error methods. Therefore, advanced cross-disciplinary techniques must be introduced to establish AIE property prediction models and enhance the development efficiency of organic AIE fluorescent materials. Machine learning (ML) is a powerful tool that has been widely applied across diverse fields to accelerate material development through complex relationship mapping of large data sets. This study presents an ML-based approach aimed at accelerating the development of organic AIE fluorescent materials. We assembled a dataset of 3,074 molecules with AIE/ACQ properties, developed a prediction model using the LightGBM ensemble learning algorithm, and used combined molecular fingerprints as input. The model achieved a remarkable independent test-set accuracy of up to 0.974 and its extrapolation ability was validated by out-of-sample validation (accuracy = 0.963). In addition, experimental validation was performed to assess the reliability of the ML prediction model in an unknown molecular space, showing a satisfactory agreement between the model predictions and the experimental results. Our work highlights the potential of ML for rapidly and accurately predicting the AIE property of unknown molecules and screening AIE materials, thereby accelerating the development of organic AIE fluorescent materials. [Display omitted]