Page 110 - Read Online

Page 12 of 15 Niu et al. J. Mater. Inf. 2025, 5, 45 https://dx.doi.org/10.20517/jmi.2025.22

Table 2. Comparison of luminescent properties of Mol1 with other synthesized similar molecules
Molecule λ (nm) a Φ QY σ (nm) f (S1 → S0) SA score
emi
emi
400 88.6% 49.8 2.00 2.0

(Mol1)
361 85.0% 41.1 1.87 2.3

(Mol2)
538 11.0% 119.9 1.66 2.7

(Mol3)
SA: Synthetic accessibility; GCN: graph convolutional neural network.

broadening. Additionally, the vibrational coupling in its high-frequency region is significantly suppressed,
resulting in a relatively narrow spectral bandwidth, which further contributes to an advantage in FWHM. In
conclusion, Mol1 demonstrates computational advantages in both PLQY and FWHM, providing theoretical
support for its exceptional photoluminescent performance within the same molecular family.

These properties suggest that Mol3, despite its structural similarity, does not meet the performance
standards expected of high-efficiency luminescent molecules. The large FWHM and low quantum yield
indicate substantial non-radiative losses, which is a critical disadvantage for any potential application in
optoelectronic devices. In contrast, Mol1 demonstrates versatility in the optical performance comparison
among these similar molecules, excelling in multiple luminescent indicators. Mol1 not only exhibits a lower
SA score of 2.0, indicating better synthetic accessibility, but also demonstrates superior luminescent
performance in terms of PLQY and FWHM. Mol1, as a proof of concept, confirms that LumiGen
successfully achieves its goal of transitioning from independent-property domains to comprehensive
performance enhancement. Its ability to combine superior photophysical properties illustrates the
effectiveness of the approach and further establishes LumiGen as a powerful framework for future
luminescent molecule design.

To verify the feasibility in the case of insufficient material data, we applied LumiGen to train a generation
model in a smaller molecular dataset of AIE materials. ASBase currently contains over 1,000 AIE functional
molecular materials, encompassing the photophysical and physicochemical properties in Supplementary
[26]
Table 5 . The Molecular Generator adeptly filled the gaps between the ASbase and the target spaces in
Supplementary Figures 12-14. Compared to the DB , the UMAP dimensionality reduction plot using high-
exp
quality AIE molecules for transfer learning shows more clustering of the sampled molecules, probably due
to the increased rigidity of AIE molecules. A notable finding was the particularly low Fsp on templates with
3
high quantum yield, suggesting targeted enhancements in molecular design during the transfer learning
process. Our comparative analysis, supported by TD-DFT calculations, revealed superior luminescent
properties in the MolElite, characterized by narrower FWHM, higher k , and higher extinction coefficient,
r
while the MolMediocrity showed less favorable characteristics in Supplementary Figures 7 and 8.
Importantly, for the molecules shown in Figure 6D and their molecular scaffold derivatives, an EQE of
[36]
nearly 25% was achieved for the sky-blue phosphorescent OLED . The adaptability for smaller datasets and
accommodating experimental variability of LumiGen highlights its practicality for real-world applications.
The emergence of sophisticated data extraction tools and advanced language models (such as ChatGPT)