An Optical Satellite Controller Based on Diffractive Deep Neural Network

Formulated as an optimal control problem, space relative trajectory planning is crucial for on-orbit servicing spacecraft on various missions. While a variety of deep-neural-network (DNN) methods have been proposed to solve the problem, they are energy-consuming and computationally consuming, which limits their on-board deployments. In this work, we proposed a kind of diffractive-deep-neural-network-based optical satellite controller, and transformed the solution of the optimal control problem into a light-field-based fine-grained regression task. Firstly, an electro-optic conversion module was designed to convert numerical relative state variables from electronic signals into light field as the input of a diffractive modulation (DM) module, the diffractive masks of which could be trained to implement complex light field transformation. We used another optic-electro conversion module to convert the light field at the output plane of DM module into electronic signals. Then, we trained the DM module to make the decoded electrical signals consistent with the desired optimal control commands. Therefore, when the light carrying input information and propagating through the well-trained diffractive masks, the DM module could perform diffract-based solution of optimal control problem. The simulation results substantiate the feasibility and effectiveness of our OC-Nets, which can achieve comparable performance to the latest classic DNN methods, except for a few acceptable errors. Different from classic models with much too energy consumption, once fabricated physically, the device of our optical controller can provide optimal control commands at the speed of light, with fairly little computational and energy consumption, and enable the on-board deployment on spacecraft.