Hybrid Neural and Deconvolution Approach for Finite-Source Reflector Design

We present a hybrid method for reflector design with finite light sources, combining a neural-network-based solver with a deconvolution-inspired iterative correction scheme. Our approach addresses the limitations of classical techniques, which often assume idealized point or parallel sources, by solving a simplified problem using a neural network and refining the solution via feedback from ray-traced simulations of the full finite-source system. We demonstrate the effectiveness of our method on a representative example, showing improved convergence toward a prescribed far-field intensity distribution compared to the approximate problem’s solution.