High Power Laser Science and Engineering, Volume. 13, Issue 3, 03000e39(2025)
Machine learning phase control of filled-aperture coherent beam combining: principle and numerical demonstration
Figures & Tables(12)
Fig. 1. System setup of deep learning phase control for filled-aperture CBC.
Fig. 4. Predicted phase versus true phase for samples of different initial RMS residual phases: (a) 0.7 rad, (b) 1.2 rad, (c) 1.8 rad and (d) 2.4 rad.
Fig. 5. Prediction error as a function of true phase: (a) cos-sin loss and two-layer output and (b) traditional MSE loss and one-layer output.
Fig. 6. System state variation during delay control process: (a) PIB of the tiled-aperture combined beam and (b) normalized intensity of the filled-aperture combined beam.
Fig. 8. Single-step residual phase for filled-aperture CBC and combining efficiency for tiled-aperture CBC with respect to training epochs.
Fig. 9. Filled-aperture CBC with dynamic phase noise: (a) time-dependent phase noise, (b) combining efficiency in open and closed loops, (c) time convergence detail from the open to the closed loop and (d) phase noise spectra in open and closed loops.
Fig. 10. Filled-aperture CBC of 36 channels with dynamic phase noise. Phase control by deep learning (a) with and (b) without strategies.
Table 1. Procedure for delay control.
View tableView in ArticleTable 1. Procedure for delay control.
Require: Parameters Tp, Td, γ and σ are defined.1: for step = 1 to ∞do 2: capture an image and feed into neural network 3: update phases φ ←φ –φ pred4: calculate delay control step s = step·Tp/(Td/3) 5: if (s modulo 3) = 1then 6: generate random vector δτ of variance σ2 7: apply positive perturbation τ + δτ 8: get PD signal as metric J+ = IPD (τ + δ τ )9: else if (s modulo 3) = 2then 10: apply negative perturbation τ – δτ 11: get PD signal as metric J– = IPD ( τ – δτ )12: else 13: calculate metric change: ΔJ = (J+ – J–)/(J+ + J–) 14: update delays τ ←τ + γδτ ΔJ / σ215: end if 16: end for
Table 2. Residual phase after one-step phase control for different channel numbers.
View tableView in ArticleTable 2. Residual phase after one-step phase control for different channel numbers.
N 9 16 25 36 Dataset size 18,000 32,000 50,000 72,000 Training epochs 90 160 250 360 Training time (h) 0.60 1.86 4.52 9.37 One-step time (ms) 1.1 1.1 1.0 1.0 φres of this network λ/70 λ/43 λ/27 λ/21 φres of traditional network λ/12 λ/4 λ/4 λ/4
Tools
Get Citation
Copy Citation Text
Hongbing Zhou, Rumao Tao, Xi Feng, Haoyu Zhang, Min Li, Xiong Xin, Yuyang Peng, Honghuan Lin, Jianjun Wang, Lixin Yan, Feng Jing. Machine learning phase control of filled-aperture coherent beam combining: principle and numerical demonstration[J]. High Power Laser Science and Engineering, 2025, 13(3): 03000e39
Paper Information
Category: Research Articles
Received: Dec. 8, 2024
Accepted: Mar. 3, 2025
Posted: Mar. 4, 2025
Published Online: Jun. 24, 2025
The Author Email: Rumao Tao (supertaozhi@163.com)
CSTR:32185.14.hpl.2025.24
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20