Transformer models for optical systems design

Transformer models offer a new approach to optical systems design by enabling broad, generative exploration of lens architectures beyond the abilities of traditional methods. Transformers can provide global optimizations as a starting point for simulated annealing, local optimization, and related techniques. Transformers capture long-range dependencies between surfaces, materials, and system-level constraints, making them well-suited for dynamic design problems such as movable lenses, tolerancing, and thermal variation. Combining supervised and reinforcement learning results in a more effective search compared to other machine learning approaches. Optimized Optics presents a Transformer-based model that searches for possible optical designs, changing surface shapes, thicknesses and glass types, just as language models choose words based on short- and long-term contexts. We apply a two-step approach to design two achromatic zoom systems. The first step applies Geometric Deep Learning to paraxial approximations to determine lens power and distances, and the second step chooses glasses, curvatures and aspheric parameters. These results demonstrate how Transformer models can complement human designers by accelerating concept generation and improving overall design-space search.