A materials informatics driven fine-tuning of triazine-based electron-transport layer for organic light-emitting devices

Materials informatics in the development of organic light-emitting diode (OLED) related materials have been performed and exhibited the effectiveness for finding promising compounds with a desired property. However, the molecular structure optimization of the promising compounds through the conventional approach, namely the fine-tuning of molecules, still involves a significant amount of trial and error. This is because it is challenging to endow a single molecule with all the properties required for practical applications. The present work focused on fine-tuning triazine-based electron-transport materials using machine learning (ML) techniques. The prediction models based on localized datasets containing only triazine derivatives showed high prediction accuracy. The descriptors from density functional theory calculations enhanced the prediction of the glass transition temperature. The proposed multistep virtual screening approach extracted the promising triazine derivatives with the coexistence of higher electron mobility and glass transition temperature. Nine selected triazine compounds from 3,670,000 of the initial search space were synthesized and used as the electron transport layer for practical OLED devices. Their observed properties matched the predicted properties, and they enhanced the current efficiency and lifetime of the device. This paper provides a successful model for the ML assisted fine-tuning that effectively accelerates the development of practical materials.


P. S3
Electron injection layer: Liq was deposited for 1 nm.
Cathode: After a striped metal mask was arranged to be orthogonal to the ITO stripe, Mg/Ag alloy (10/1, wt/wt) was deposited for 80 nm with 0.5 nm sec −1 rate.Then, Ag was deposited for 20 nm with 0.2 nm sec −1 rate.
For OLED (Figure S1), Hole injection layer: the mixture containing 50 wt% of HIL and 50 wt% of HTL was co-deposited for 10 nm with 0.15 nm sec −1 rate.
Hole transport layer: HTL was deposited for 10 nm with 0.15 nm sec −1 rate.
Electron blocking layer: EBL was deposited for 10 nm with 0.15 nm sec −1 rate.
Emitting layer: the mixture containing 5 wt% of Emitter and 95 wt% of Host was co-deposited for 25 nm with 0.18 nm sec −1 rate.
Hole blocking layer: HBL was deposited for 5 nm with 0.15 nm sec −1 rate.
Electron transport layer: the mixture containing 50 wt% of the triazine derivatives and 50 wt% Liq was codeposited for 25 nm with 0.15 nm sec −1 rate.
Electron injection layer: Yb was deposited for 2 nm with 0.10 nm sec −1 rate.
Cathode: After a striped metal mask was arranged to be orthogonal to the ITO stripe, Mg/Ag alloy (10/1, wt/wt) was deposited for 80 nm with 0.5 nm sec −1 rate.Then, Ag was deposited for 20 nm with 0.2 nm sec −1 rate.
After the deposition, the obtained assembly of multi-layers was encapsulated with a glass cap and ultraviolet ray-curable epoxy resin (purchased from Nagase Chemtex).The encapsulation was conducted in a nitrogen atmosphere having an oxygen-and-moisture content of below 1 ppm within a glove box.
To the suspension was added potassium phosphate aqueous solution (2 mol L −1 , 4.8 mL) and the mixture was stirred at reflux temperature for 16 hours.After cooling to the room temperature, water and methanol were added to the mixture.After the crude product was corrected by filtration, it was dissolved to toluene and the toluene solution was stirred with activated carbon powder at 100 °C.Then, the clear colorless filtrate was obtained through filtration by using cerite bed.The crude product was purified by recrystallization from toluene to give 2,4-bis(biphenyl-4-yl)-6-[2,4-(pyridine-2-yl)-biphenyl-4'-yl]-1,3,5-triazine (T2-7191) as white solid (1.25 g, 1.81 mmol, 57 %).
The random grid search algorithm was used for the hyperparameter optimization.Although the boosting-tree type models showed higher scores initially, they resulted in overfitting of training data.In the tuning process of the boosting-tree type algorithms, MAE of training dataset became close to zero, but MAE of test data did not decrease through the further optimization.The behavior seems to be derived from the small amount of the training dataset.

P. S16
Comparison of Tg prediction models (Table S1)

P. S17
Examples of the structure-predicted properties in the last screening step (Table S2) Table S2.The triazine compounds excluded in the last screening step: molecular structures, its calculated LUMO levels, predicted Tg(2D), Tg(3D), μe, and the reason not to be chosen as the potentially practical ETL materials.

P. S18
Luminescence spectra and other properties of the fabricated OLED devices (Figure S4) The slight difference in the range from 470 nm to 500 nm seems to be derived from the change of carrier balance.
Table S3 Properties of the fabricated OLED devices at 10 mA cm −2 of current density.

Figure S1
Figure S1 Materials for blue light emission OLED used in the present work.

Figure S2
Figure S2 Data distribution of the initial datasets.(a) electron mobility estimated from the EOD

Figure S4 .
Figure S4.Emission spectra of the OLED device fabricated in this work.

Table S1
Properties of the previously reported Tg prediction models.The prediction model based on 1944 literature data was applied to 40 OLED-related compounds.‡TheTg(2D)andTg(3D) models were trained on the same train/test separation.The MAE and R 2 values are slightly different from Table1due to the change of the train/test separation.