Prompt-to-prescription: towards generative design of diffraction-limited refractive optics

Abstract

The design of high-performance optical systems remains a specialized domain gated by the limited availability of expert engineers, creating a bottleneck that stalls innovation despite the growing demand for imaging hardware. While deep learning has improved parameter optimization, it has yet to address the fundamental challenge of conceptualizing valid optical architectures from functional requirements. Here, we present an end-to-end generative framework that couples the semantic reasoning of Large Language Models (LLMs) with a differentiable ray-tracing engine to democratize the synthesis of diffraction-limited optical prescriptions. By treating optical design as a semantic-to-physical translation task, the system autonomously interprets prompts ranging from high-level end-user requests to rigorous technical specifications. We demonstrate the framework's versatility across three distinct regimes: (1) finite-conjugate industrial metrology systems, where the model autonomously enforces application-specific constraints such as telecentricity to achieve diffraction-limited performance; (2) a suite of infrared objectives (NIR, SWIR, and LWIR), demonstrating the framework's ability to synthesize valid topologies and optical prescriptions for non-visible spectral bands, and (3) complex aspheric mobile lenses, where the system successfully navigates the high-dimensional optimization landscape to produce high-resolution designs suitable for modern sensors. Validated against industry-standard simulation tools, these results establish a new paradigm for automated optical engineering, bridging the gap between semantic intent and physical realization.