Tree-Based Machine Learning Model for Fluorescence Lifetime Prediction in Organic Compounds
Journal of Fluorescence · 19.06.2026 · 36 (6) · 4121-4139
DOI: 10.1007/s10895-026-04839-9 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Fluorescence, online erschienen am 19.06.2026. Verfasst von Dragoș-Cătălin Vovea und 1 weiteren Autor:innen. Beteiligte Einrichtung: Faculty of Physics, Babeș-Bolyai University, Str. M. Kogălniceanu 1, Cluj-Napoca, RO- 400084, Romania. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 29 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Fluorescence lifetime is a key photophysical property that reflects the excited-state decay dynamics of organic chromophores and is essential for applications in fluorescence imaging, sensing, optoelectronics, and materials screening. In this study, a tree-based machine-learning framework was developed to predict fluorescence lifetimes using chromophore–solvent data derived from the Deep4Chem database. Molecular structures were encoded through RDKit descriptors, Morgan and MACCS fingerprints, solvent descriptors, engineered interaction features, and photophysical variables, enabling structural, electronic, geometrical, and environmental effects to be captured in a unified descriptor space. Several regression algorithms were benchmarked, and LightGBM was selected as the final model owing to its balance of predictive accuracy, computational efficiency, scalability, and interpretability. The optimized model was evaluated using cluster-stratified 10-fold cross-validation, achieving a mean MAE of 0.8324 ± 0.0617 ns and a mean R² of 0.7523 ± 0.0317. Error analysis showed that the model performs most reliably in the intermediate lifetime range, where data density is highest, while larger deviations occur for sub-nanosecond and long-lifetime chromophores due to relative-error amplification, experimental uncertainty, and limited sample representation. External validation against TD-DFT benchmarks and independent experimental data further demonstrated that the model captures global fluorescence lifetime trends and can serve as a rapid screening tool rather than a replacement for detailed spectroscopy or quantum-chemical calculations. SHAP analysis revealed that fluorescence lifetime is mainly governed by quantum yield, molecular geometry, emission-related descriptors, topological indices, and fragment-level structural motifs, indicating that the model learns physically meaningful relationships between molecular structure, solvent environment, excited-state stabilization, and non-radiative relaxation.
Schlagwörter
Stichwörter der Autor:innen
Autor:innen und Institutionen
- Faculty of Physics, Babeș-Bolyai University, Str. M. Kogălniceanu 1, Cluj-Napoca, RO- 400084, Romania.
- Faculty of Physics, Babeș-Bolyai University, Str. M. Kogălniceanu 1, Cluj-Napoca, RO- 400084, Romania. vasile.chis@ubbcluj.ro.
Zitieren
APA
Vovea, D. C., & Chiș, V. (2026). Tree-Based Machine Learning Model for Fluorescence Lifetime Prediction in Organic Compounds. Journal of Fluorescence, 36(6), 4121-4139. https://doi.org/10.1007/s10895-026-04839-9
Vancouver
Vovea DC, Chiș V. Tree-Based Machine Learning Model for Fluorescence Lifetime Prediction in Organic Compounds. Journal of Fluorescence. 2026;36(6):4121-4139. doi:10.1007/s10895-026-04839-9
BibTeX
@article{Vovea2026,
author = {Vovea, Dragoș-Cătălin and Chiș, Vasile},
title = {Tree-Based Machine Learning Model for Fluorescence Lifetime Prediction in Organic Compounds},
journal = {Journal of Fluorescence},
year = {2026},
volume = {36},
number = {6},
pages = {4121--4139},
doi = {10.1007/s10895-026-04839-9}
}
EUVIMED