Differentiable Ray Tracing for End-to-End Optical Designs and Modeling

Traditionally, the design and modeling of optical systems have heavily relied on accurately calculating ray intersections through a physically accurate simulation of light transport within an optical system. However, it is also crucial to understand how selected merit functions evolve concerning component parameters. Mathematically, this requires differentiability, involving the computation of derivatives for an optical system. Emerging as a technique, differentiable ray tracing facilitates the tracing of rays and their variations concerning parameters of interest. The application of differentiability streamlines the solution of inverse problems, akin to the training of neural networks, enabling direct optimization through gradient descent. Differentiability also facilitates the simultaneous optimization of front-end optical system modeling (hardware) and back-end image processing algorithms (software), such as neural networks, making end-to-end hardware-software joint design feasible from a numerical standpoint. The first half of this webinar covers the basic principles of differentiable ray tracing. We will delve into its applications in various areas, including freeform and lens designs, sensitivity analysis, AI-driven designs (or deep lens), and its broader utility in addressing inverse problems within optical system modeling. The second half of this webinar will cover more advanced recent developments of differentiable ray tracing in automatic lens design, end-to-end designs, and more. Future perspectives on end-to-end designs will also be presented.