Frontier molecular orbital weighted model based networks for revealing organic delayed fluorescence efficiency

Organic light-emitting diodes (OLEDs) have become a cornerstone in display and lighting technologies, driving the demand for efficient organic thermally activated delayed fluorescence (TADF) materials. Modeling and predicting the complex luminescent properties of TADF materials, particularly their photoluminescence quantum yield (PLQY), requires a deep understanding of molecular electronic structures. Currently, most deep learning methods for predicting molecular properties rely heavily on geometric structures, often neglecting the critical electronic structures that determine the luminescent process. This limitation makes it difficult to meet the demands of large-scale material screening in the OLED field.

 

To address this challenge, a team of scientists, led by Professor Yue Wang from Jilin University and Professor Hai Bi from Jiuhua Laboratory, China, has developed the Electronic Structure-Infused Network (ESIN), a novel deep learning model that integrates molecular geometry and electronic structure data. This work is published in Light: Science & Applications. ESIN leverages the principles of frontier molecular orbitals (FMOs) to enhance the model's predictive accuracy and interpretability. By focusing on the HOMO (highest occupied molecular orbital), LUMO (lowest unoccupied molecular orbital), HOMO-1, and LUMO+1 orbitals, ESIN can identify key molecular features that influence PLQY, as well as the interactions between these electronic orbitals, providing valuable insights into the luminescent process.

 

One of the key innovations of ESIN is its sampling method focused on D-A (donor-acceptor) structures. A local sampling approach involves selecting fixed numbers of sampling centers based on ring structures to capture topologies that maintain complete chemical properties. By focusing on key positions and their electronic structural relationships, the sampling method ensures that the model can handle TADF molecules of varying sizes effectively. ESIN then extracts the geometric and electronic structure representations of these molecular orbitals, enabling efficient characterization of TADF molecules and accurate assessment of their luminescent properties. This approach enables the model to quickly screen highly efficient TADF materials even with small datasets.

 

The research team collected a dataset under conditions where TADF data is scarce, see Figure 2. The team also ensured that the dataset covered a wide range of molecular structures and properties to test the robustness of the ESIN model. Using this dataset, ablation studies were conducted, and it was found that incorporating electronic structure characterization into the model significantly improved prediction accuracy, enabling more efficient screening of TADF materials.

 

Moreover, the model introduces attention weights for key molecular orbitals. These attention weights offer a clear visualization of how different electronic structure parts of the molecule contribute to the overall PLQY. For instance, the model has shown that the HOMO and LUMO primarily determine the PLQY of TADF materials. Additionally, the combined effects of HOMO, LUMO, HOMO-1, and LUMO+1 influence the luminescent process. For specific TADF systems, the distribution of attention weights exhibits clear patterns. Based on the fitting results of these attention weight distributions, adjustments to the functional groups corresponding to the frontier molecular orbitals can be made to design the luminescent properties of the target system. This insight is crucial for designing new TADF molecules with optimized electronic structures.

 

"For designing a new efficient molecule, we typically start with a reported core framework and explore different combinations and modifications to optimize the electronic distribution. The trained ESIN model, which can filter out molecules with lower PLQY, provides an effective method for optimizing design. For a target D-A structure, the first step involves analyzing the relationship between attention weights and PLQY to fit distribution patterns. Next, by identifying the substructures that have the greatest impact on the distribution patterns, we can focus on the key substructures. Finally, through model predictions and experimental feedback, we refine the distribution pattern model to find the best modification sites, thus rapidly identifying efficient TADF materials," the researchers summarized.

 

Two new diphenylquinoxaline-based molecules, DPQ-DPAC and DPQCN-DPAC, were synthesized and experimentally tested. Modifying the acceptor with a cyano group in DPQCN-DPAC increased the attention weight of the LUMO+1 and reduced the variance of attention weights, enhancing the PLQY value. This validation confirms ESIN's ability to evaluate the luminescent potential of TADF molecules by investigating molecular structure-dependent FMO characteristics.

 

In summary, this finding underscores the importance of integrating both geometric and electronic structure information in the model, highlighting the superior performance of ESIN in predicting PLQY and optimizing TADF material design.

---

---