Multiagent Data Mining for Enhanced Machine Learning Prediction of Delayed Fluorescence Materials.

One of the primary bottlenecks hindering the widespread adoption of AI-driven exploration of novel materials in the field of organic light-emitting diodes (OLEDs) is the labor-intensive process of constructing high-quality data sets required for training accurate property prediction models. Consequently, the limited availability of complete experimental data impedes a deeper understanding of the underlying materials' mechanisms. In this work, we introduce a multiagent artificial intelligence framework that autonomously extracts information from PDF-formatted literature, enabling accurate retrieval of molecular structures, properties, detailed measurement conditions, and OLED device architectures. This is achieved by decoupling the complex task of data extraction from PDFs and integrating easily accessible, multiple specialized AI agents. Using this framework, we curated and publicly released data from over 1600 open-access publications, covering most known thermally activated delayed fluorescence (TADF) materials. When applied to this real-world data set, the trained machine learning (ML) models in our framework demonstrated high accuracy in predicting six key properties essential for materials screening. We further investigated the influence of additional experimental conditions on the predictive performance of the machine learning models. Our results demonstrate that lowering the barrier to acquiring OLED-related data using our framework can significantly enhance the prediction accuracy of ML models for emitter screening, accelerating AI-driven material discovery in this field.