Integrating AI into optical engineering: experiences in workflow transformation

Abstract

We report experiences integrating a large-language-model agent into a production camera-lens design workflow. Rather than claiming the AI designs better lenses, we describe a division of labor: the AI writes analytical tooling and runs trade-space sweeps that would otherwise consume weeks of expert time, while the engineer defines problems, verifies physics, and decides. The most consequential product is a parallelized as-built tolerance engine, 1,024 Monte-Carlo trials in approximately three minutes versus an estimated two-day serial baseline, that makes manufacturing yield cheap enough to evaluate routinely during design exploration. That engine now extends across object distance (as-built yield at near conjugates and autofocus-versus-fixed-focus) and the system accumulates a durable knowledge library, indexed for semantic retrieval, so the agent reuses hard-won fixes instead of re-deriving them. We describe the system architecture, the workflow boundary between human and machine, and what did not work: a controlled study showing no significant strategy-selection effect, and a case where the agent pursued a flawed approach for hours until one expert sentence redirected it. A second engineer independently reproduced the tooling-partner workflow for stray-light analysis, and for feeding a downstream image-simulation pipeline, suggesting the method generalizes beyond one person and one domain. This continues a 26-year thread: a 2000 optimization macro, an expert system that produced the top two designs for the 2002 IODC Lens Design Problem, and now an AI agent, of encoding design knowledge to constrain the optimizer. Results are from a single wide-angle mobile-camera design class; broader validation is ongoing.