The Predictions of Optoelectronic Attributes of LED
5.2 Neural Network and Its Learning Algorithm
In this Chapter, a three-layer neural network is mainly taken as the prediction model. The network consists of one input layer, one hidden layer, and one output layer. Each layer is composed of several nodes. Fig. 5.1 shows the basic structure of a three-layer neural model used in our research. The sigmoid function is used as the node’s transfer function in the neural model. The error back-propagation (BP) algorithm is the learning rule adopted for network’s
training process in all studies.
Fig. 5.1 The architecture of neural network model.
5.2 Simulations
In this research, 293 data provided by Epitech Technology Corp. are used for study and simulation. To demonstrate the feasibility of the neural models, three data sets, named Set-1, Set-2 and Set-3 were randomly reorganized from the original data. For each data set, 196 data were used for training and 97 data were used for testing. The mean absolute error (MAE) and mean absolute percentage error (MAPE) are treated as the measurs for the performances of all models. Table 1 shows the example of data simulated.
Table 5.1 The example of data simulated.
Recipe DBR (20mA) Iv20 (100mA)Iv100 Wde Vf20 Vf100 Iv Wdc Vf
2M-GM21E001 3 1.03 4.07 569.5 1.849 1.935 42.5 571.5 2.06
2M-GM21E003 3 0.30 1.05 563.8 1.862 1.947 18.5 565.8 2.07
2M-RM01E001 2 0.54 1.44 629.2 1.736 1.832 40.0 627.1 1.98
2M-RM02S209 3 2.19 7.64 620.4 1.731 1.828 136.0 621.0 2.00
2M-YM01E009 2 1.36 4.00 586.3 1.848 1.938 91.3 589.5 2.04
2M-YM01S001 2 1.23 3.62 585.5 1.850 1.943 90.0 588.7 2.08
2M-YM02S014 3 2.38 8.07 586.3 1.843 1.929 165.4 592.1 2.03
Three optoelectronic attributes of LED chip, including wavelength, intensity and forward voltage, were predicted. The detailed simulations are presented as follows.
The prediction of wavelength:
In this study, a size of 6-5-1 neural model is used for all experiments. The learning rate
= 0.1 and momentum ζ = 0.3 are adopted in neural model. The inputs and output are listed as,
Inputs: Recipe, DBR and Wde Output: Wdc
In the inputs, all recipes were encoded by using four-digital numbers. The coding numbers are shown in Table 2. DBR is the type of distributed Bragg reflector. Wde is the test value of wavelength by EL in the epitaxy growth stage. In this study, the predicted MAEs of Set-1, Set-2 and Set-3 are 0.2792, 0.2830 and 0.31222, respectively. Their MAPEs are 0.047%, 0.048% and 0.052%. Fig. 5.2, Fig. 5.3 and Fig. 5.4 depict the superposition curves of the actual and predicted wavelengths for these three data sets.
Table 5.2 The recipe codes
2M-GM21E001 0001
2M-GM21E003 0010
2M-RM01E001 0011
2M-RM02S209 0100
2M-RM02S212 0101
2M-RM02S214 0110
2M-RM02S215 0111
2M-YM01E009 1000
2M-YM01S001 1001
2M-YM02S014 1010
2M-YM02S015 1011
Fig. 5.2 The superposition curves of the actual and predicted Wavelengths of Set-1.
(solid line: actual values, dotted line: predicted values)
Fig. 5.3 The superposition curves of the actual and predicted Wavelengths of Set-2.
(solid line: actual values, dotted line: predicted values)
Fig. 5.4 The superposition curves of the actual and predicted Wavelengths of Set-3.
(solid line: actual values, dotted line: predicted values)
The prediction of intensity:
In this study, a size of 8-5-1 neural model was used for all experiments. The learning rate
=0.2 and momentum ζ=0.3 are adopted in the model. The inputs and output are listed as,Inputs: Recipe, DBR, Iv20, Iv100 and Wde
Output: Iv
In the inputs, Iν20 and Iν100 are testing intensities by using currents 20 mA and 100 mA, respectively. Both values were obtained in the epitaxy growth stage also. In this study, the predicted MAEs of Set-1, Set-2 and Set-3 are 4.0598, 4.0120 and 4.2340, respectively. Their MAPEs are 3.94%, 3.68% and 3.96%. Fig. 5.5, Fig. 5.6 and Fig. 5.7 plot the superposition curves of the actual and predicted intensities for these three data sets.
Fig. 5.5 The superposition curves of the actual and predicted Intensities of Set-1.
(solid line: actual values, dotted line: predicted values)
Fig. 5.6 The superposition curves of the actual and predicted Intensities of Set-2.
(solid line: actual values, dotted line: predicted values)
Fig. 5.7 The superposition curves of the actual and predicted Intensities of Set-3.
(solid line: actual values, dotted line: predicted values)
The prediction of forward voltage:
In this study, a size of 7-5-1 neural model was used for all experiments. The learning rate
=0.2 and momentum ζ=0.3 were adopted. The inputs and output of neural model are listed as,
Inputs: Recipe, Wde, Vf20 and Vf100
Output: Vf
In the inputs, Vf20 and Vf100 are testing forward voltages by using currents 20 mA and 100 mA, respectively. Both values were also obtained in the epitaxy growth stage. In this study, the predicted MAEs of Set-1, Set-2 and Set-3 are 0.0089, 0.0092 and 0.0106, respectively. Their MAPEs are 0.44%, 0.45% and 0.52%. Fig. 5.8, Fig. 5.9 and Fig. 5.10 show the superposition curves of the actual and predicted intensities for these three data sets.
Voltage
Fig. 5.8 The superposition curves of the actual and predicted Voltage of Set-1.
(solid line: actual values, dotted line: predicted values)
Voltage
Fig. 5.9 The superposition curves of the actual and predicted Voltage of Set-2.
(solid line: actual values, dotted line: predicted values)
Voltage
Fig. 5.10 The superposition curves of the actual and predicted Voltage of Set-3.
(solid line: actual values, dotted line: predicted values)
5.3 Conclusion
In this research, the optoelectronic attributes of LED chip, including the luminous intensity, wavelength and forward voltage, were predicted by using neural network techniques.
From the simulation results, we could find that all predictions are quiet accurate and within their individual tolerance. In other words, the predicted results can be used as the references to help engineer to adjust the parameters in the epitaxy growth process of LED, if the prediction result shows some abnormal phenomena. Undoubtedly, the mechanism proposed in this study could also be used as a useful training tool for those junior technicians who have no enough experience in the LED manufacturing process. Therefore, we conclude that the neural network prediction mechanism not only can greatly increase the fabrication efficiency in the real-line LED manufacturing operation, but also can reduce the production cost for the
company.
However, only the traditional neural model was applied in this research. It might not have the best research result. In our future research, other neural model and learning algorithm will be continuously studied in this field.