Wavefront prediction with reservoir computing for minimizing the effects of angular anisoplanatism

Area of Science:

• Astronomy
• Astrophysics
• Machine Learning

Background:

• Recurrent neural networks estimate atmospheric turbulence effects on natural guide star images.
• Current methods are effective for low turbulence but limited for moderate-turbulence multilayer models.
• Enhancements in network parameter optimization show potential for broader applicability.

Purpose of the Study:

• To propose spatio-temporal learning using reservoir computing (echo state networks) for estimating spatially variant point spread functions (PSFs).
• To enable improved image restoration of point-source exo-atmospheric objects outside the isoplanatic patch.
• To extend turbulence estimation methods to moderate-turbulence multilayer models.

Main Methods:

• Utilized reservoir computing, specifically echo state networks, for discriminative learning.
• Modeled the forward problem by training a reservoir computer with time-series perturbations from multiple natural guide stars.
• Incorporated known site profile data to optimize the model for training and testing.

Main Results:

• The proposed echo state network method effectively estimates spatially variant PSFs.
• Demonstrated potential for improved image restoration of exo-atmospheric objects.
• Showcased the ability to handle aberrations over a wide, anisoplanatic field.

Conclusions:

• Echo state networks offer a promising approach for estimating atmospheric turbulence effects.
• This method enhances image restoration capabilities for astronomical observations.
• The technique is suitable for moderate-turbulence multilayer models and anisoplanatic conditions.