Solving inverse freeform illumination problems via deep learning

Designing freeform optics for finite-étendue sources is challenging, as most methods assume ideal, zero-étendue sources. While this simplifies calculations, the performance degrades with real sources. We introduce a deep learning method to predict freeform surfaces directly. The trained model learns the entire solution space, enabling prediction of freeform surfaces for unseen source-target combinations. This represents the first generalizable deep learning framework for nonimaging freeform design and is a first step toward data-driven methods for finite-étendue illumination systems.