Using Diffraction Deep Neural Networks for Indirect Phase Recovery Based on Zernike Polynomials

Fang Yuan, Yang Sun, Yuting Han, Hairong Chu, Tianxiang Ma, Honghai Shen2024 · DOI

TL;DR

This paper introduces a novel indirect phase recovery method using Diffractive Deep Neural Networks (DDNNs) to reconstruct wavefront aberrations represented by Zernike polynomials. This approach offers a potentially faster and more cost-effective alternative to traditional wavefront sensors for adaptive optics systems.

What Problem Does This Paper Address?

Accurately measuring and correcting wavefront distortions is crucial for the performance of optical systems, particularly in adaptive optics (AO). Traditional methods, like Shack-Hartmann sensors, are precise but can be expensive and computationally intensive. While deep learning has shown promise, existing neural network approaches often require significant computational resources and can be slower than hardware-based solutions. This paper aims to bridge the gap by proposing an indirect phase retrieval method using DDNNs, which leverage optical principles for computation, to reconstruct Zernike polynomial coefficients from intensity measurements.

Methods

The core of this work is the application of a Diffractive Deep Neural Network (DDNN) for indirect phase recovery. Unlike traditional neural networks that rely on electronic computation, DDNNs utilize a series of diffractive layers to perform computations optically.

Here's a breakdown of their approach:

  • Indirect Phase Recovery: Instead of directly measuring the phase, the DDNN learns a mapping from the distorted input beam's intensity to the coefficients of Zernike polynomials that describe the wavefront aberration.
  • Diffractive Deep Neural Network (DDNN): The DDNN is constructed from multiple diffractive layers. Each layer contains numerous "diffraction units" (pixels) that act like neurons. The complex-valued transmission coefficients of these units are trainable parameters.
  • Optical Computation: Light propagates through these layers, undergoing modulation (amplitude and phase) at each diffractive element. This process mimics the operations of a neural network, with the interaction between light and matter performing the computation.
  • Zernike Polynomials: The method specifically targets Zernike polynomials, which are a standard set of orthogonal polynomials used to describe common optical aberrations. The DDNN is trained to output coefficients corresponding to these polynomials.
  • Sigmoid Transformation: To represent both positive and negative Zernike coefficients in the output intensity (which is always non-negative), a sigmoid transformation is applied to the coefficients before they are mapped to intensity values in specific regions of the output plane.
  • Post-processing: After the DDNN processes the input, the resulting intensity distribution in the output plane is captured. Simple post-processing, including summing intensities and applying an inverse sigmoid transformation, is used to extract the predicted Zernike coefficients. These coefficients are then used to reconstruct the phase.
  • Training: The DDNN is trained using an Adam optimizer to minimize a combined loss function that includes both the intensity error (to avoid stray light) and the coefficient accuracy error.

Datasets

The authors created two distinct datasets for training and validation:

  • Dataset 1: Consists of phase distortions generated from single Zernike polynomials (Z1 to Z10). It includes 10,000 training images and 2,000 testing images, with each polynomial represented by 1000 training and 200 testing images.
  • Dataset 2: Features phase distortions generated by combining Zernike polynomials of orders one to ten. This dataset comprises 12,000 images (10,000 for training, 2,000 for testing).

The wavefront errors in these datasets were generated within a specific range, mimicking typical astronomical observation scenarios, with peak-to-valley (PV) values between 0.3λ and 3λ, and an average RMS error of 0.25λ.

Key Insights and Ideas

  • Optical Computation for Phase Recovery: The most significant insight is the successful application of DDNNs for indirect phase recovery. This demonstrates the potential of using optical hardware to perform complex machine learning tasks, offering advantages in speed and power consumption.
  • Zernike Coefficient Mapping: The DDNN effectively learns a complex mapping between distorted wavefronts and their corresponding Zernike polynomial coefficients, which are then encoded in the output intensity distribution.
  • Indirect vs. Direct Phase Recovery: By focusing on indirect recovery (predicting coefficients rather than the phase directly), the method simplifies the output requirements and leverages the inherent properties of Zernike polynomials.
  • Generalization Capabilities: The study investigates the network's ability to generalize to unseen data and varying input phase magnitudes, showing promising results within a certain range.
  • Layer Depth Impact: The research explores how the number of diffractive layers affects performance, concluding that a five-layer network offers a good balance between accuracy and generalization.

Key Results

  • Average RMS Error: The simulation results show an average root mean square (RMS) error of 0.086λ for phase inversion, meeting imaging quality requirements.
  • Single-Order Recognition: The DDNN successfully identifies and distinguishes between individual Zernike aberrations (Figure 10).
  • Combined Phase Recovery: The network can recover combined aberrations from the first ten Zernike orders with significant reduction in RMS error (Figure 13).
  • Non-Degenerate Response: The network exhibits a non-degenerate response, meaning it can differentiate between input phases with the same amplitude but opposite signs, producing distinct outputs.
  • Generalization Performance: The five-layer network demonstrates good generalization, with RMS errors controlled around 0.1λ for input-phase RMS sizes up to 0.75λ.

Limitations and Future Work

  • High-Order Aberrations: The paper acknowledges that the complexity of the recovered wavefront is currently limited by the designed output intensity distribution. This restricts its effectiveness for recovering high-order complex phases.
  • Physical Implementation: While the paper presents simulation results, the physical implementation of the diffractive layers and their performance in real-world optical experiments are areas for future work.
  • Range Limitations: The network's accuracy is not precisely guaranteed for phase distributions beyond its training range. Further research is needed to extend this range.
  • Cost and Complexity: While promising, the paper notes that further research is required to explore aspects such as the physical implementation and performance enhancement of this scheme.

Summary updated

BibTeX
@misc{yuan2024using,
  author = {Fang Yuan and Yang Sun and Yuting Han and Hairong Chu and Tianxiang Ma and Honghai Shen},
  title = {{Using Diffraction Deep Neural Networks for Indirect Phase Recovery Based on Zernike Polynomials}},
  year = {2024},
  doi = {10.3390/s24020698},
}
Figure