Predicting freeform surface topologies for illumination applications via deep learning (Invited Paper)

In illumination optics, the goal is to modify a light source’s spatial distribution to achieve a specific irradiance target. By using freeform surfaces, the emitted light can be transformed into arbitrary irradiance patterns. Despite significant advances in freeform optics design, mostly for ideal light sources, current methods offer limited control over the surface shape, generally resulting in globally convex or concave surfaces. In an illumination context, smooth and oscillating freeform surfaces are possibly more interesting, but calculation methods are currently non-existent. This presentation introduces a deep learning approach for predicting such complex freeform topologies, capable of rapidly generating optics that transform a prescribed light source into arbitrary irradiance patterns. This allows for the creation of surfaces with convex, concave, and saddle regions, showcasing the potential of deep learning in accelerating illumination design.