Estimating differential pistons for the Extremely Large Telescope using focal plane imaging and a residual network

TL;DR

This paper presents a deep learning method using a ResNet architecture to estimate differential piston aberrations in the ELT from Shack-Hartmann wavefront sensor images, improving image quality control.

Contribution

It introduces a novel neural network approach for differential piston estimation from SH-WFS images, demonstrating robustness under various observing conditions.

Findings

Deep learning can accurately estimate differential piston from SH-WFS images.

Temporal averaging enhances piston signal detection.

Performance is resilient to polychromaticity and detector noise.

Abstract

As the Extremely Large Telescope (ELT) approaches operational status, optimising its imaging performance is critical. A differential piston, arising from either the adaptive optics (AO) control loop, thermomechanical effects, or other sources, significantly degrades the image quality and is detrimental to the telescope's overall performance. In a numerical simulation set-up, we propose a method for estimating the differential piston between the petals of the ELT's M4 mirror using images from a 2x2 Shack-Hartmann wavefront sensor (SH-WFS), commonly used in the ELT's tomographic AO mode. We aim to identify the limitations of this approach by evaluating its sensitivity to various observing conditions and sources of noise. Using a deep learning model based on a ResNet architecture, we trained a neural network (NN) on simulated datasets to estimate the differential piston. We assessed the robustness of the method under various conditions, including variations in Strehl ratio, polychromaticity, and detector noise. The performance was quantified using the root mean square error (RMSE) of the estimated differential piston aberration. This method demonstrates the ability to extract differential piston information from 2x2 SH-WFS images. Temporal averaging of frames makes the differential piston signal emerge from the turbulence-induced speckle field and leads to a significant improvement in the RMSE calculation. As expected, better seeing conditions result in improved accuracy. Polychromaticity only degrades the performance by less than 5% compared to the monochromatic case. In a realistic scenario, detector noise is not a limiting factor, as the primary limitation rather arises from the need for sufficient speckle averaging. The network was also shown to be applicable to input images other than the 2x2 SH-WFS data.