Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

From orbital analysis to active learning: an integrated strategy for the accelerated design of TADF emitters

Jean-Pierre Tchapet Njafa, Steve Cabrel Teguia Kouam, Patrick Mvoto Kongo, Serge Guy Nana Engo

Abstract

Thermally Activated Delayed Fluorescence (TADF) emitters must satisfy two competing requirements: small singlet-triplet energy gaps for thermal upconversion and sufficient spin-orbit coupling for fast reverse intersystem crossing. Predicting these properties accurately demands expensive calculations. We address this using a validated semi-empirical protocol (GFN2-xTB geometries, sTDA/sTD-DFT-xTB excited states) on 747 molecules, combined with charge-transfer descriptors from Natural Transition Orbital analysis. The hole-electron spatial overlap She emerges as a key predictor, accounting for 21% of feature importance for the triplet state alone. Our best model (Support Vector Regression) reaches MAE = 0.024 eV and R2 = 0.96 for $ΔE_{ST}$. Active learning reduces the data needed to reach target accuracy by approximately 25% compared to random sampling. Three application domains are explored: NIR-emitting probes for bioimaging, photocatalytic sensitizers, and fast-response materials for photodetection.

From orbital analysis to active learning: an integrated strategy for the accelerated design of TADF emitters

Abstract

Thermally Activated Delayed Fluorescence (TADF) emitters must satisfy two competing requirements: small singlet-triplet energy gaps for thermal upconversion and sufficient spin-orbit coupling for fast reverse intersystem crossing. Predicting these properties accurately demands expensive calculations. We address this using a validated semi-empirical protocol (GFN2-xTB geometries, sTDA/sTD-DFT-xTB excited states) on 747 molecules, combined with charge-transfer descriptors from Natural Transition Orbital analysis. The hole-electron spatial overlap She emerges as a key predictor, accounting for 21% of feature importance for the triplet state alone. Our best model (Support Vector Regression) reaches MAE = 0.024 eV and R2 = 0.96 for . Active learning reduces the data needed to reach target accuracy by approximately 25% compared to random sampling. Three application domains are explored: NIR-emitting probes for bioimaging, photocatalytic sensitizers, and fast-response materials for photodetection.

Paper Structure

This paper contains 19 sections, 2 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Theoretical and computational framework
  3. Computational details
  4. Methodological benchmarking
  5. High-throughput protocol and dataset
  6. High-fidelity reference calculations
  7. Machine learning methodology
  8. Active learning strategy
  9. Results and discussion
  10. Model interpretability via SHAP analysis
  11. Mechanistic rationalization
  12. Active learning performance
  13. Acquisition function comparison
  14. Model performance
  15. Application-specific molecular design
  16. ...and 4 more sections

Figures (4)

  • Figure 1: SHAP feature importance analysis for $\Delta E_{\text{ST}}$ prediction. Energy features contribute $\sim$57%, CT descriptors $\sim$34%, with $S_{he}^{T_1}$ (21%) being the dominant CT feature.
  • Figure 2: Active learning vs. random sampling learning curves for $\Delta E_{\text{ST}}$ prediction. AL (blue) consistently outperforms random sampling (red). Shaded regions indicate $\pm 1$ standard deviation.
  • Figure 3: Comparison of acquisition functions. Hybrid and Diversity strategies outperform random sampling, while UCB shows significant degradation.
  • Figure 4: SVR model performance for $\Delta E_{\text{ST}}$ prediction with $R^2 = 0.960$ and MAE $= 0.024$ eV.