Optical Design Benchmarks
Two AI capabilities of the etendue engine, each benchmarked on real optical specs. Starting-Point Quality measures the seeds the engine generates: the starting design it produces for a spec. Engine Performance measures the full agentic designer that turns a spec into a finished, optimised lens, starting from a seed of its own.
The full agentic design: the engine takes a spec, generates a seed, then optimises it end-to-end into a finished lens. Each point is one completed design, scored on two axes (optical Perfection and Production readiness) and placed by its distance to target. With starting points shown, the arrow traces its journey from the seed to the optimised result; closer to the centre is closer to the target. Click any point to inspect its design and download it (.etd / .zmx).
What the numbers mean
Each finished design is scored on two axes, and the set Score is their average. Optical proximity (P) is how close the design lands to the optical target (RMS spot size and Strehl versus the specification), where 100 means the target is met. Realism (R) is how practical the design is to actually build. It averages the realism factors we measure for each design: manufacturability (tolerances, edge and centre thickness, clear apertures), standard-glass availability, element economy (element count versus a typical design), and topology sanity (no near-flat surfaces wasting an element). Tolerance yield is measured separately, and until it is evaluated it is left out of the score rather than assumed, so R only ever reflects what we actually measured.
How good the engine's seeds are, before any optimisation. For each spec it builds a starting design from a classic lens archetype (doublet, Cooke, retrofocus…); each is scored by how close it already sits to the diffraction limit, i.e. its distance to the Airy radius, and how much geometric room it leaves an optimiser.