Atmospheric Turbulence Phase Reconstruction via Deep Learning Wavefront Sensing

Yutao Liu, Mingwei Zheng, Xingqi Wang2024 · DOI

TL;DR

This paper introduces a novel deep learning approach using a U-Net architecture to reconstruct atmospheric turbulence-induced phase distortions in free-space optical communication. By training on intensity images, the method achieves rapid and accurate phase reconstruction, outperforming traditional methods in both speed and accuracy, and demonstrating strong generalization across various turbulence levels.

What Problem Does This Paper Address?

Free-space coherent optical communication offers high data rates but is severely hampered by atmospheric turbulence. This turbulence causes random fluctuations in the refractive index of the atmosphere, leading to distortions in the amplitude and phase of the transmitted light beam. These distortions degrade the signal quality, increasing the bit error rate. Accurately reconstructing the turbulence-induced phase distortion is crucial for compensating these effects and enabling reliable communication. Traditional methods for phase reconstruction, such as those relying on wavefront sensors or iterative phase inversion algorithms, suffer from slow convergence, inadequate accuracy, or limitations in handling scintillation effects. This paper aims to address these limitations by proposing a faster, more accurate, and robust method for atmospheric turbulence phase reconstruction.

Methods

The core of this paper is a deep learning-based approach utilizing a U-Net convolutional neural network. The U-Net architecture, known for its effectiveness in image segmentation tasks, is adapted here for phase reconstruction. The network takes a single-channel 2D image of a Gaussian beam affected by atmospheric turbulence as input. The U-Net consists of a contracting path (downsampling stages) for feature extraction and an expanding path (upsampling stages) for precise localization and reconstruction. Key features of their U-Net implementation include:

  • Padding: Used in convolutional layers to maintain image size during downsampling, preventing information loss.
  • Skip Connections: Horizontal arrows in the U-Net diagram represent concatenation of feature maps from the contracting path to the expanding path, allowing the network to retain fine-grained details.
  • Downsampling: Achieved using 2x2 max pooling operations.
  • Upsampling: Performed using 2x2 transpose-convolution operations.

The network is trained to predict the turbulence phase from the distorted intensity image. The loss function used is a combination of binary cross-entropy (for predicting phase values) and Mean Squared Error (MSE) to evaluate the accuracy of the reconstructed phase.

Datasets

The authors generated a dataset for training and testing their U-Net model. This dataset consists of simulated atmospheric turbulence phase screens and corresponding distorted Gaussian beam intensity images. They simulated turbulence under three distinct intensity levels, quantified by the atmospheric refractive index structure constant ($C_n^2$):

  • Weak turbulence: $C_n^2 = 5 \times 10^{-17} m^{-2/3}$
  • Moderate turbulence: $C_n^2 = 5 \times 10^{-15} m^{-2/3}$
  • Strong turbulence: $C_n^2 = 1 \times 10^{-13} m^{-2/3}$

For training, they used 6000 samples, and for testing, 2400 samples were utilized. The paper also mentions an experimental setup used to gather empirical data, which contributed to the dataset for validation.

Key Insights and Ideas

  • Intensity-to-Phase Reconstruction: The most significant insight is the successful demonstration that a deep learning model can directly reconstruct the complex phase distortion from a simple intensity image, bypassing the need for explicit wavefront sensors.
  • U-Net's Suitability: The U-Net architecture proves highly effective due to its ability to capture both high-level contextual information (through downsampling) and precise spatial details (through upsampling and skip connections), which are crucial for accurate phase reconstruction.
  • Generalization Across Turbulence Levels: The model exhibits remarkable generalization capabilities, performing consistently well across weak, moderate, and strong turbulence conditions without requiring separate models for each. This is evidenced by the similar loss function curves and MSE values across different $C_n^2$ levels.
  • Speed Advantage: The deep learning approach offers a substantial speedup in phase reconstruction compared to traditional methods, making it suitable for real-time applications in dynamic atmospheric conditions.

Key Results

  • Reconstruction Accuracy: The proposed U-Net model achieved an average loss function value of 0.00027 post-convergence and a mean squared error (MSE) of 0.0003 for individual turbulence reconstructions during simulation. Experimental validation yielded an MSE of 0.0007.
  • Reconstruction Time: The model demonstrates a rapid average reconstruction time of approximately 0.14 seconds, significantly faster than CNN and AlexNet models which took 0.31 s and 0.29 s respectively. This is more than a 50% reduction in time.
  • Convergence Speed: The U-Net model converges much faster than traditional deep learning models like CNN and AlexNet, achieving a low loss function value within 2500 iterations, whereas the other models still showed a downward trend.
  • Robustness: The model's performance remained stable across different turbulence intensities, indicating robustness against varying atmospheric conditions.

Limitations and Future Work

  • Limitations:
    • The paper notes that while the model performs well, there are still "burrs" or fluctuations in the loss function curves during training, indicating that the reconstructed phase might not always be closer to the actual phase than the previous iteration.
  • The experimental results show some discrepancies, particularly in regions with significant phase variations, suggesting that the model can be challenged by extreme phase fluctuations and noise.
  • Future Work:
    • The authors suggest that the proposed method benefits the application of free-space coherent optical communication, implying that future work could focus on integrating this phase reconstruction technique into actual communication systems for real-time compensation.
  • Further improvements could involve exploring more advanced deep learning architectures or training strategies to enhance accuracy in extreme turbulence scenarios and further reduce reconstruction time.

Summary updated

BibTeX
@misc{liu2024atmospheric,
  author = {Yutao Liu and Mingwei Zheng and Xingqi Wang},
  title = {{Atmospheric Turbulence Phase Reconstruction via Deep Learning Wavefront Sensing}},
  year = {2024},
  doi = {10.3390/s24144604},
}
Figure