Physics-driven self-supervised learning for non-modulated pyramid wavefront sensing

Area of Science:

• Astronomy
• Optical Engineering
• Machine Learning

Background:

• Non-modulated pyramid wave-front sensors (PWFS) offer high sensitivity for extreme adaptive optics (ExAO) but exhibit significant nonlinearity.
• Existing deep learning approaches require large labeled datasets and lack physical interpretability.

Purpose of the Study:

• To introduce PINN-Pyr, a novel physics-informed neural network for self-supervised wavefront reconstruction.
• To overcome the limitations of conventional methods in terms of nonlinearity, data requirements, and interpretability.

Main Methods:

• PINN-Pyr embeds a differentiable forward optical model within a U-Net architecture.
• The network performs self-supervised learning, mapping intensity patterns to Zernike coefficients without paired training data.

Main Results:

• PINN-Pyr extends the linear dynamic range of PWFS through its nonlinear mapping.
• It achieves lower residual root-mean-square (RMS) error and superior Strehl ratio (SR) stability compared to MVM and data-driven methods.
• Performance is robust under strong turbulence and low signal-to-noise ratio (SNR) conditions.

Conclusions:

• PINN-Pyr offers a robust and efficient solution for wavefront reconstruction in extreme adaptive optics.
• This physics-constrained deep learning approach is suitable for next-generation large-aperture telescopes.