Direct neural prediction of freeform illumination optics

Designing freeform optics for extended light sources remains a challenge in illumination design, since conventional design methods for zero-étendue sources are difficult to extend to general, finite-étendue sources. As an alternative to these conventional methods, this work introduces a framework for direct prediction of zero-étendue freeform illumination surfaces. A multi-stage training strategy is presented for a deep neural network enabling near-instant prediction of smooth freeform geometries for point sources and random target irradiance distributions of varying sizes. The predicted designs achieve high irradiance fidelity and serve as effective initialization for differentiable fine-tuning, requiring only a couple of optimization iterations to reach ultra-precise irradiance control. This represents the first, to the best of our knowledge, deep learning framework for nonimaging freeform illumination design with zero-étendue sources. While this work focuses on zero-étendue sources, the multi-configuration capability of the framework provides a fundamental base that can be extended to eventually capture the full spatial and angular emission characteristics of finite-étendue sources.