Network training parameters exploration for generating starting points of freeform reflective imaging system design based on machine learning

Abstract

Using freeform optical surface in an imaging optical system is a revolution in the field of optical design. Introducing machine learning into freeform imaging optical design will significantly reduce the human effort and even beginners in optical design will be able to perform difficult design tasks. Machine learning has been successfully applied to the immediate generation of starting points with various system specifications for the design of freeform reflective imaging systems. However, the parameters used in the network training, which are the key points in the whole design framework, are determined without proper guidance, which may significantly affect the actual performance of the networks. In this paper, a comprehensive exploration of the training parameters of the neural network used for starting points generation of freeform reflective systems is conducted. The parameters include the number of layers, the number of nodes in the layers, the type of activation function, the type of loss function, the type of optimization algorithm, and the value of learning rate. A detailed comparison and analysis of different training parameters are demonstrated on the training results and the imaging performance of validation output systems with large amount of random system specifications input. Using the obtained results designers can choose proper parameters accordingly and get desired neural networks with shorter training time and better performance. The results also offer insight in the design of imaging systems with other system configurations and more advanced system specifications.