Understanding the\nKey Factors for Photoinduced Radical\nGeneration in Crystalline Triphenylamines Using Experiment and Machine\nLearning

In this study, we combine experiments, calculated properties,\nand\nmachine learning (ML) to design new triphenylamine-based (TPA) molecules\nthat have a high photoinduced radical (PIR) generation in crystals.\nA data set of 34 crystal structures was extracted from the Cambridge\nCrystallographic Data Centre. Eighteen structures with experimentally\nreported PIR values from 0 to 0.85% were used to build an ML model\ntrained using Random Forest that achieves an average leave-one-out\ntest set error of 0.173% PIR. The ML model was used to screen the\nremaining 16 compounds, of which 4 were selected and subsequently\ncompared with the experimentally measured PIR%. The predicted PIR%\ndemonstrated good agreement with the measured values of TPA bis-urea\nmacrocycle host–guest complexes and non-macrocyclic compounds\nof TPAs. Examining a broad set of molecular architectures/scaffolds\nallows for investigating the structural and electronic properties\nthat lead to high PIR generation. We found very different trends for\nmacrocycles, linear TPAs, and mono TPAs, where mono TPAs consistently\nhave the lowest PIR generation. Macrocycles tend to have the highest\nPIR generation, especially for systems with benzene and fluorobenzene\nguests. Although linear analogs overall perform worse than macrocycles,\nthey display clear trends with increasing excited-state dipole moment,\noscillator strength, and electron–hole covariance, while decreasing\nionization potential and interatomic distance are generally correlated\nwith higher PIRs. What is consistently observed is that higher PIRs\nare seen for brominated analogs. Our study, therefore, provides guidelines\nfor future design strategies of TPAs for PIR generation.