Regulation-aware freeform headlamp reflector design with differentiable ray tracing

Headlamp design is an essential aspect of the automotive industry, often relying on reflector systems composed of freeform surfaces. Traditional methods depend on manually adjusting surfaces to ensure reflected rays meet regulation requirements, which is time-consuming and often limited by design features. In this work, we propose a regulation-aware approach to freeform headlamp reflector design using differentiable ray tracing. The reflectors are modeled with B-spline surface arrays, and we optimize the control point grid through a customized ray tracing pipeline utilizing triangle mesh acceleration structures for efficient ray-surface intersections. Additionally, we use multi-layer perceptrons (MLPs) to generate light intensity distributions guided by loss functions based on regulation constraints. Experimental results show that this neural process generates the required four types of passing-beam and driving-beam distributions defined in the ECE R112. Using these distributions as targets, we can optimize reflector shapes in about 200 iterations. We also export the optimized reflectors as CAD models, achieving 2.5× faster simulation in our pipeline with accuracy comparable to the commerical optical design software, aka., LightTools.