Deep learning-based focal plane wavefront sensing for exoplanet imaging

Exoplanet imaging presents significant challenges due to the extreme contrast and close angular separation between exoplanets and their host stars. Although coronagraphs and adaptive optics largely address these issues, high-contrast imaging instruments are still limited by residual wavefront errors, primarily from instrumental phase aberrations. Focal-plane wavefront sensing is appropriate to handle pupil-plane phase errors because it measures the aberrations directly at the science focal plane. The current sensing approaches can nonetheless be slow, unstable, and often designed for specific instruments.

In this thesis, new deep learning-based methods are explored to provide a fast, robust, and flexible solution for focal-plane wavefront sensing that can be applied to various instrumental designs. State-of-the-art deep convolutional neural networks are implemented and trained in a supervised way, achieving high performance and showing good robustness to changing aberration regimes and noise content. The problem of phase retrieval behind vortex phase masks is notably revisited with deep neural networks and simulated data. A new CNN-based approach that achieves a 100\% science duty cycle using only focal-plane images is developed and tested with both scalar and vector vortex coronagraphs.

Supervised learning techniques require ground truth phase aberrations for training the neural networks, that cannot be obtained accurately in real systems. To address this limitation, an autoencoder architecture is proposed, that is composed of a convolutional neural network as the encoder and a differentiable simulator as the decoder. As a result, the models are trained exclusively on observed images, enforcing the latent space to represent phase aberrations. Additionally, a variational component is introduced to predict probability distributions and assess phase uncertainties.

The deep learning-based methods developed in this thesis are successfully tested in-laboratory using the Subaru Coronagraphic Extreme Adaptive Optics (SCExAO) instrument. Pre-training on large simulated datasets enhances the performance on small in-lab sets, and closed-loop experiments demonstrate robust convergence, effectively correcting introduced aberrations. The simulator-based autoencoder, tested across different wavefront aberration regimes, delivers very good PSF reconstructions and phase estimations. Validating the method with a simple optical propagation model shows great promise, and further developments of the simulator should improve its robustness against the varying conditions occurring in real systems.

This thesis demonstrates that methods based on deep artificial neural networks can offer an accurate, fast and robust solution for focal plane wavefront sensing, paving the way for future developments and on-sky applications.