TL;DR
This paper introduces SimAE, a novel simulator-based autoencoder for focal-plane wavefront sensing. By integrating a differentiable optical simulator into a deep learning architecture, SimAE can learn to correct instrumental aberrations without requiring ground-truth phase data, making it a promising unsupervised approach for exoplanet imaging.
What Problem Does This Paper Address?
High-contrast imaging of exoplanets, crucial for detecting and characterizing these distant worlds, is severely hampered by instrumental aberrations. These aberrations manifest as quasi-static speckles in the science images, obscuring the faint light from the exoplanet. Focal-plane wavefront sensing (FPWFS) methods are designed to measure and correct these aberrations by analyzing the point spread function (PSF) in the focal plane. While deep learning, specifically Convolutional Neural Networks (CNNs), has shown great promise for FPWFS, traditional supervised learning approaches require extensive labeled datasets (i.e., known aberration phases), which are difficult to obtain in real-world astronomical observations. This paper addresses the challenge of developing an effective FPWFS method that can learn from data without explicit labels, leveraging the physics of the optical system.
Methods
The core innovation is a simulator-based autoencoder (SimAE) architecture. This architecture consists of two main components:
- Encoder: A deep CNN (specifically EfficientNet-B4) that takes a PSF as input and outputs Zernike coefficients, which represent the pupil-plane phase aberrations.
- Decoder: A differentiable optical simulator that takes the predicted Zernike coefficients and reconstructs a noiseless PSF.
This setup allows for an unsupervised learning approach. The autoencoder is trained to reconstruct the input PSF. The loss function is specifically designed for photon-limited data, using the Poisson distribution's log-likelihood, rather than a standard mean-square error. This approach constrains the encoder to predict physically meaningful Zernike coefficients that, when fed into the simulator, reproduce the observed PSF. The authors also explore transfer learning by fine-tuning a pre-trained SimAE model on a single observed image, allowing for rapid adaptation to new data.
Datasets
The study utilizes simulated datasets generated to mimic astronomical observations. These datasets consist of pairs of input PSFs and corresponding aberration phase maps. The simulations cover two aberration regimes:
- Low aberration: 70 nm root-mean-square (rms) wavefront error (WFE) over 20 Zernike modes.
- High aberration: 350 nm rms WFE over 100 Zernike modes.
The PSFs are generated using the PROPER optical propagation package. The data also includes variations in signal-to-noise ratio (SNR) due to photon noise and simulated atmospheric turbulence residuals. No new benchmark datasets are introduced; rather, the authors generate controlled simulations to test their method against various conditions.
Key Insights and Ideas
- Physics-informed deep learning for unsupervised learning: By incorporating a physical simulator as the decoder, the SimAE can learn to infer aberration parameters without ground truth labels. This is a significant advantage for applications where obtaining labeled data is challenging.
- Differentiable simulators as powerful decoders: The use of a differentiable optical simulator allows the entire network to be trained end-to-end, enabling the encoder to learn parameters that are directly interpretable within the physical model.
- Robustness to noise and aberrations: The SimAE demonstrates comparable performance to supervised CNNs across various SNR levels and aberration strengths, suggesting its viability for real-world scenarios.
- Effective transfer learning: The ability to fine-tune a pre-trained SimAE on a single new image significantly improves performance, especially for data that deviates from the original training distribution. This highlights the model's adaptability and potential for rapid deployment on-sky.
Key Results
- Performance: The SimAE achieves performance nearly identical to a standard supervised CNN for reconstructing Zernike coefficients, even in high aberration regimes (350 nm rms over 100 modes).
- SNR robustness: Both SimAE and CNN models show similar performance across a range of SNRs, with SimAE exhibiting a slight advantage at very low SNRs due to its specialized loss function.
- Turbulence residuals: When trained on data with atmospheric turbulence residuals (which the decoder does not explicitly model), the SimAE shows only a "reasonable degradation" in performance compared to a turbulence-free case, outperforming a standard CNN in this scenario.
- Transfer learning: Fine-tuning a SimAE on a single test sample leads to substantial performance improvements, particularly for data with higher WFE or lower SNR than the original training data. This fine-tuning process is computationally efficient, taking about 20 seconds per sample on a GPU.
Limitations and Future Work
- Decoder limitations: The current simulator is a simplified model. For real-world applications, a more complete simulator incorporating elements like coronagraphs and other instrumental components will be necessary.
- Turbulence handling: While the SimAE shows some robustness to atmospheric turbulence, explicitly incorporating this information into the decoder (e.g., using wavefront sensor telemetry) could further improve performance.
- Variational inference: The authors suggest exploring variational inference to predict posterior distributions of Zernike coefficients, which could enhance model robustness by incorporating prior knowledge.
- On-sky application: The ultimate goal is to apply this method to real astronomical data, which will require further validation and adaptation to the complexities of actual observations.
Summary updated