Author Affiliations
11. School of Optoelectronics Engineering, Xi'an Technological University, Xi'an 710021, China22. Xi'an Institute of Applied Optics, Xi'an 710065, Chinashow less
Fig. 1. One via wavelength grating structure diagram
Fig. 2. Neural network structure
(a): Forward simulation network; (b): Reverse-design network
Fig. 3. (a) Forward simulation Loss function curve; (b) Inverse design Loss function curve
Fig. 4. Series neural network
R: Expected spectral response; R': Forward simulation prediction spectrum; D: Sample structure of the original training set; D': Reverse design forecast structure. The red frame is the forward simulation network. The Loss function is modified to solve the problem that the network cannot be fitted due to the non-uniqueness of the data. The middle layer is the output of reverse design and the input of forward simulation
Fig. 5. Series network loss function curve
Fig. 6. Red, green and blue are the spectral response curves reported by references, and black curves are
Fig. 7. RCWA numerical simulation curves with inverse design of series network
Black curve is target spectrum with a reflectivity of 100%; red-green-blue curves are RCWA simulation curves of reverse design with the reflectivity of 98.91%, 99.98% and 99.88% at the peak wavelengthes of 479.5, 551.0 and 607.0 nm, respectively
| 网络层数 | 消耗时间/s | 均方误差 |
---|
Model_A | 5 | 无法收敛 | 0.513 905 | Model_B | 4 | 465 | 0.014 | Model_C | 3 | 382 | 0.10 |
|
Table 1. Evaluation indexes of network with different hidden layers
| 网络结构 | 消耗时间/s | 均方误差 |
---|
Model_1 | 50, 200, 200, 50 | 91.64 | 0.068 427 | Model_2 | 50, 200, 200, 200 | 121.89 | 0.082 540 | Model_3 | 50, 200, 500, 200 | 190.29 | 0.031 505 | Model_4 | 50, 200, 500, 500 | 373.53 | 0.033 413 |
|
Table 2. Evaluation indexes of network with different network structures
| 样本数 | 消耗时间/s | 均方误差 |
---|
Batch_size_1 | 32 | 3 483 | 0.025 68 | Batch_size_2 | 64 | 574 | 0.034 15 | Batch_size_3 | 128 | 381 | 0.028 23 | Batch_size_4 | 256 | 282 | 0.031 52 | Batch_size_5 | 512 | 268 | 0.040 17 |
|
Table 3. Evaluation indexes of network with different Batch sizes
| Method | H1 /nm | H2/ nm | F | Λ /nm | N |
---|
1 | RCWA | 90 | 150 | 0.5 | 360 | 2.2 | Network | 88.441 5 | 150.476 | 0.485 823 | 362.866 6 | 2.193 37 | 2 | RCWA | 70 | 110 | 0.7 | 360 | 2.1 | Network | 67.619 93 | 120.378 0 | 0.676 11 | 360.900 5 | 2.069 56 | 3 | RCWA | 90 | 90 | 0.65 | 360 | 2.3 | Network | 86.095 81 | 104.919 36 | 0.620 73 | 350.569 | 2.302 507 |
|
Table 4. Comparison of structural parameters
| H1 /nm | H2 /nm | F | Λ /nm | N |
---|
Red (607 nm) | 60.5 | 104.5 | 0.46 | 363.6 | 2.04(Si3N4) | Green(551 nm) | 59.2 | 105.8 | 0.46 | 323.5 | 2.04(Si3N4) | Blue(479.5 nm) | 58.5 | 106.5 | 0.46 | 273.15 | 2.04(Si3N4) |
|
Table 5. Red, green and blue structural parameters
相关性 | 无相关 | 弱相关 | 中度相关 | 强相关 |
---|
r | r<0.1 | 0.1<r<0.3 | 0.3<r<0.5 | 0.5<r<1 |
|
Table 6. Evaluation Index of correlation coefficient
| H1 /nm | H2 /nm | F | Λ/ nm | N | r |
---|
Red (607 nm) | 79.011 7 | 120.795 | 0.584 | 381.602 | 1.926 | 0.685 1 | Green(551 nm) | 53.246 52 | 107.609 | 0.411 | 326.287 | 2.052 | 0.813 4 | Blue(479.5 nm) | 45.680 69 | 93.037 | 0.602 | 283.814 | 2.014 | 0.789 6 |
|
Table 7. Reverse design parameters