ВИМІРЮВАННЯ НЕЙРОМЕРЕЖЕЮ ХВИЛЬОВОГО ФРОНТУ НА ОСНОВІ ФОРМИ СВІТЛОВОГО ВІДБИТКУ

Determining the wavefront is necessary for identifying and compensating for aberrations in various optical systems: from telescopes to the human eye. Specialized devices use different sensors for this purpose, such as the Hartmann-Shack sensor, the pyramid wavefront sensor, and others. Wavefront sensors have limitations; they can introduce distortions due to glare formation and also increase the size of optical systems. To overcome these obstacles, modern researchers are working with neural networks, which have been a focus of the scientific community in recent years. One direction of this work is the use of reinforcement learning methods, which have several advantages over feedforward neural networks, particularly the ability to adjust their actions, which is relevant for measurement systems. For example, in, a system was built that uses a flexible mirror to reproduce the assumptions of a neural network agent regarding the values of Zernike coefficients that describe the wavefront, followed by determining the quality of aberration compensation. This work proposes a solution based on Proximal Policy Optimization (PPO) with a simplified optical system consisting of only two detectors and one beam splitter, and the agent’s assumptions are reproduced through virtual ray tracing. Based on the coordinates of intersections points of rays and detectors, the binary pixel image is built using calculation of A-shape with the following pixelation by the Bresenham algorithm. The quality of the result is measured as a Jaccard distance between real and generated light spots. As a result of performance testing, an accuracy of 95% was achieved in 80 steps for wavefronts described by three lower-order Zernike coefficients. Future improvements are planned by involving more advanced reinforcement learning methods.