TURBO-RL: turbulence mitigation using reinforcement learning for severe optical aberrations

Optical distortion or aberration remains a vital challenge that prohibits high-resolution imaging in various applications such as space domain awareness, terrestrial remote sensing, and astronomy. However, due to the stochastic nature of these optical distortions, reducing their effect without directly measuring wavefronts is challenging. Furthermore, in the case of extreme turbulence, due to the limited size of the lenslet array in the wavefront sensor, the sensor fails to correctly quantify or minimize the image distortions of a guide star from turbulence. While numerous studies have shown effectiveness of guide star-based adaptive optics in mitigating mild turbulence, severe turbulence has remained a persistent challenge. To target this, we present TURBO-RL: TURBulence mitigatiOn using Reinforcement Learning, which uses just a single optical element (e.g., deformable mirror) to estimate and correct the wavefront errors from a guide star. TURBO-RL adopts reinforcement learning with a convolutional neural network to extract and estimate turbulence. Unlike other methods, TURBO-RL is capable of guide star imaging in severe turbulence (D/r₀=100) with only about 590 photons, making it possible to overcome the strong turbulence and possibly replace bulky and expensive wavefront sensors.