Nonimaging Optical Design with Supporting Quadric Method and Deep Neural Network

Background: The supporting quadric method (SQM) is a versatile method for designing freeform optics for desired irradiance redistribution, but the time of solution optimization increases rapidly with the refinement of the mapping grid. Objective: As the complexity of light distribution is getting higher and higher, time consumed will also increase exponentially. This paper proposes an idea of applying the deep neural network method to optical design. Methods: In this article, we established a special corresponding relationship and prepared a dataset, which underwent deep network learning and training. Finally, a hybrid design method of deep learning and optical design was realized and verified. Results: Compared with the traditional method, this method is more efficient. Here, we used a deep neural network(DNN) to accelerate the freeform optical design. After the DNN was trained by a sample set consisting of a uniform pattern and eight different Chinese characters represented by an array with 11 × 11, it can generate a character's reflector within few milliseconds. Conclusion: As proof of this new method, a character pattern reflector was manufactured and tested, and the experimental irradiance distribution was found close to the expectation, which means that the neural network has the excellent capability to memorize all of the learned characters. SQM combined with DNN has the potential to establish a particular “optical font library” and even offers a promising path for rapid freeform optical design to realize the function of “optical typography”.