| Thermally activated delayed fluorescence（TADF）materials,as the third-generation organic light-emitting diode（OLED）materials,have attracted much attention because they are free of noble metals and can theoretically achieve 100%internal quantum efficiency（IQE）.However,the high-precision theoretical calculations and experimental development of materials are complicated,time-consuming and expensive,it is a big challenge to predict properties and screen high-performing materials facilely and accurately.Given this,this paper employs quantum chemical calculations combined with machine learning methods to investigate several hot issues of TADF materials,including the prediction of intersystem crossing（ISC）rate,singlet-triplet energy splitting(ΔEST)of boron-based molecules and full width at half maximum（FWHM）of fluorescence spectrum.These properties can be accurately predicted by established machine learning models with relevant descriptors,which accelerate the screening high-performance TADF materials.The main research contents of this paper are summarized as follows:（1）For TADF materials,triplet excitons can be converted to singlet excitons by efficient ISC and reverse intersystem crossing（RISC）,achieving 100%IQE theoretically.However,the facile and accurate prediction of ISC and RISC rates for a large number of TADF molecules is quite challenging.Here,two computationally affordable and high-performance（Pearson correlation coefficient=0.8-0.9）machine learning models are established based on improved descriptors.With the reliable GBRT model,nine promising TADF molecules with theoretically predicted ISC rates>7×107 s-1 are proposed by a virtual screening of 564 candidate molecules constructed from 20 unique units,which can be used for guide synthesis and device fabrication.In addition,the possible properties of the high ISC rate molecules are summarized from the perspective of statistics.The methodology would complement quantum mechanical calculations as an efficient alternative approach for the prediction of ISC rates,and provide theoretical guidance for designing high-performance TADF molecules.（2）Boron-based TADF materials have been widely used in OLEDs due to their high efficiency,stability,and ability to achieve multicolor luminescence.However,it is difficult to accurately predict theirΔEST values by using conventional quantum chemical calculations,probably due to their double excitations and multireference characters.Here,a machine learning model with lower computational cost and less errors is established based on the relevant descriptors with Root Mean Square Error（RMSE）=0.052 e V and Mean Absolute Error（MAE）=0.043 e V.Moreover,several newly reported boron-based molecules were predicted using the GBRT-KNN-KRR model,in most cases the deviation is within 0.05 e V.This finding provides an efficient and accurate method for predicting theΔEST of boron-based TADF molecules and contributes to accelerating the development of boron-based TADF materials.（3）Materials with a small FWHM exhibit high color purity,which can be widely used in high-definition display devices.However,there are few reports to predict the FWHM of organic light-emitting materials.Here,two reliable machine learning models（RMSE=9.90 nm,10.47 nm）are established for molecules with emission wavelengths in the range of 400-500 nm based on the selected descriptors.The FWHM of molecules with emission wavelengths in the range of 300-399nm and 500-599 nm were also predicted using the GBRT-KNN-SVM model.Though there is an increased error,it is basically consistent with the trend of the experimental values.The results may provide a possible method for predicting the FWHM and accelerating the discovery of narrowband TADF materials. |
