Adaptive Data Splitting Strategy to Enhance Knowledge Learning for Phosphorescence Lifetime Prediction

Abstract

Room-temperature phosphorescence (RTP) lifetime is a critical property for optoelectronic and bioimaging applications, yet its experimental measurement is labor-intensive and datasets remain scarce. Herein, we construct RTPLD, a curated dataset of 681 RTP molecules collected from extensive literature with the assistance of large language models (LLMs), providing a standardized foundation for data-driven modeling of RTP systems. To enable reliable learning from this limited but chemically diverse dataset, we propose DMFsplit, an adaptive data splitting strategy that integrates the global molecular features and local structural variations. By dynamically balancing the two-level features, DMFsplit generates more representative training sets and improves chemical knowledge acquisition in machine learning models. Extensive benchmarks across ten machine learning models demonstrate that DMFsplit consistently outperforms random and three local structure-based data splitting strategies. Coupling with the complementary ECMA feature representation, DMFsplit-XGBoost achieves R2 values of 0.84 and 0.75 on the independent and external test sets, respectively. In addition, SHAP analysis further reveals key structural motifs governing the RTP lifetime, guiding rational experimental design. Collectively, this work establishes a publicly available dataset, a generalizable data splitting strategy, and a reliable predictive model, facilitating knowledge-driven RTP research.