Freeform surface topology prediction for prescribed illumination via semi-supervised learning

Area of Science:

• Optics and Photonics
• Machine Learning Applications
• Optical Design

Background:

• Freeform optical design faces challenges in creating smooth, shallow topologies for specific illumination patterns.
• Applications include glare-free illumination and thin beam shaping elements.
• Machine learning (ML) shows promise for inverse problems in optics, but is underutilized in freeform illumination design.

Purpose of the Study:

• To present a rapid, standalone framework for predicting freeform surface topologies.
• To generate prescribed irradiance distributions from a predefined light source.
• To explore ML for complex freeform illumination design challenges.

Main Methods:

• A 2D convolutional neural network (CNN) models the relationship between target irradiance and freeform topology.
• A semi-supervised learning approach trains the network using a loss function comparing obtained and input irradiance.
• A secondary network replaces traditional Monte-Carlo raytracing for efficiency.

Main Results:

• The framework rapidly predicts smooth freeform topologies for arbitrary irradiance patterns.
• The semi-supervised learning method outperforms supervised learning, acknowledging multiple topologies can produce similar patterns.
• The developed ML model effectively addresses the inverse problem in freeform illumination design.

Conclusions:

• The presented ML framework offers an efficient solution for designing freeform optical surfaces for illumination.
• This approach enables the rapid generation of complex, smooth freeform topologies.
• The work inspires further ML applications in unsolved problems within freeform illumination design.