Prediction of Light Extraction Efficiency Using a Transfer Learning Model

Calculation of optical efficiency in an organic light emitting diode works as a bottleneck for the device geometry optimization process. However, almost all commercial simulation tools obtain solutions based on physical rules. Therefore, the calculation speed of the simulation tools is governed by the simulation space. However, prediction based on artificial neural networks is independent of the simulation space size, hence the calculation speed can be boosted while maintaining the complexity of structure which requires a large number of equations under the physical rules. [1] Despite its advantage, replacing the numerical simulation with an artificial neural network is not always efficient due to the computational load required for the generation of training data.<br/>Transfer learning is a useful method for reducing the required number of training data. There are few papers that used transfer learning to calculate the optical properties [2, 3]. However, these studies have been conducted on a case where a plane wave, which is a relatively simple case, is incident from the outside. Furthermore, the previous papers do not consider training models with data that contains errors.<br/>This paper focuses on the OLED system and presents a method to design a network even when there is not enough trainable data. In this paper, we will explain network structure design techniques in cases where experimental or theoretical data are insufficient. To apply the transfer learning, we pre-trained a 15-layers deep learning model to calculate the light extraction efficiency spectrum of a 6-parameter OLED structure with 200,000 trainable data sets.<br/>First, we demonstrate a transfer learning method by predicting the light extraction efficiency of a six-layer OLED system (7-parameter model) based on a neural network trained for predicting the light extraction efficiency of a five-layer OLED device (6-parameter model). We reuse some layers from the five-layer OLED prediction network and re-train the network for a six-layer OLED device with only 4,000 data. Generating speed for the whole model is 100 times faster than the non-transfer case. Root mean squared error (RMSE) is used to mark the accuracy of each model. The RMSE of the 6-parameter model is 0.0029 and the RMSE of the 7-parameter model reaches 0.0085 with only 2,000 data. For the 8-parameter model system, the RMSE is 0.0098 when applying the part of the 6-parameter network<br/>Second, this paper also generates the deep learning model that predicts the light extraction efficiency from the measured dataset which contains both functional error and random error. Measured data not only contains random errors but also functional errors. Deep learning models cannot predict random error intrinsically, but functional error can be predicted by the deep learning model. We apply both Gaussian random error and functional error at input data that represent the structural parameter and transfer the deep learning model that is trained with error-free data. With only 8,000 training data, the model can predict the functional error with an RMSE error of 0.005.<br/>This paper suggests a method for generating the deep neural network with a small amount of additional training data.<br/><br/>[1] Jiang, J., Chen, M. & Fan, J.A. Deep neural networks for the evaluation and design of photonic devices. Nat Rev Mater 6, 679–700 (2021). [2]. Dong Xu, Yu Luo, Jun Luo, Mingbo Pu, Yaxin Zhang, [2] Yinli Ha, and Xiangang Luo, "Efficient design of a dielectric metasurface with transfer learning and genetic algorithm," Opt. Mater. Express 11, 1852-1862 (2021)<br/>[3] Zhang, J., Qian, C., Fan, Z., Chen, J., Li, E., Jin, J., Chen, H., Heterogeneous Transfer-Learning-Enabled Diverse Metasurface Design. Adv. Optical Mater. 2022, 10, 2200748.