Optimizing branch-point detection using machine learning

In this paper, we leverage wave-optics simulations to explore a machine learning approach for identifying the locations of branch points in Shack–Hartmann wavefront sensor (SHWFS) data. We train a convolutional neural network to receive a single SHWFS sub-aperture irradiance pattern image and estimate whether the pupil of that area of interest (AOI) contains at least one branch point. We test the model alongside a branch-point detection approach that relies on the beam spread (D_4σ) of individual SHWFS irradiance patterns to determine whether or not a branch point is present within the AOI. We explore the performance of both the machine-learning- based and D_4σ-based branch-point identification approaches across a wide range of optical turbulence conditions, quantified by Rytov numbers (<i>R</i>) ranging from 0 - 2, as well as in the presence of additive noise. To quantify our performance, we use various metrics commonly used for binary classifiers, including the F₁ score and the Receiver Operating Curve (ROC). By and large, we find that the machine learning model outperformed the D_4σ approach. Implementing machine learning of this caliber into branch-point identification provides the basis for future studies exploring improved adaptive-optics compensation in regions of strong turbulence.