BRENNEN

Industries / Technology / Lasers and photonics

Optics

Optics is the work of designing and making the lenses, mirrors, prisms and coatings that bend, focus and reflect light, and of measuring their shape to fractions of a wavelength.

5

Milestones achieved, from 1872 to 2022.

3

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1872

    Microscopes designed by calculation

    Carl Zeiss opened a workshop for precision mechanics and optics in Jena in 1846 and hired the physicist Ernst Abbe in 1866. Until then microscopes were built by trial and error; from 1872 the Zeiss works designed them from Abbe's calculations, by the account of the ZEISS archives, with "vastly improved optical properties".

    zeiss.comzeiss.com

    Credit

    Ernst Abbe and Carl Zeiss

  2. 1884

    Glass made for lenses

    Paul Riedel, the first member of Abbe's scientific staff, evaluated the optical properties of glass samples made by Otto Schott from 1880 to 1885. In 1884 Abbe, Schott, Carl Zeiss and Roderich Zeiss founded the Glastechnisches Laboratorium Schott & Genossen in Jena to make optical glass. Abbe insisted that the new optical glass be sold to other makers as well, which ZEISS says lifted the whole German optics industry.

    zeiss.com

    Credit

    Otto Schott, Ernst Abbe, Carl Zeiss and Roderich Zeiss; the company is now SCHOTT

  3. 1980s

    A diamond cuts the mirror

    Lawrence Livermore National Laboratory built the Large Optics Diamond Turning Machine (LODTM) in the early 1980s to make large aspheric metal mirrors, curved in shapes that are not part of a sphere, for experimental military lasers. It takes parts up to 1.65 m across and 1,360 kg; a diamond tool cuts a reflecting surface that often needs no further polishing, while seven interferometers measure where the tool is relative to the part. The US National Bureau of Standards, now NIST, tested its positioning accuracy in the field in 1984, without cutting tests, and the machine later cut three secondary mirrors for the Keck telescopes. Livermore wrote in 2001 that, eighteen years after its first operations, measuring devices were still not good enough to confirm its accuracy.

    Lessona machine can outrun the instruments needed to check its work.

    econnection.mst.edunvlpubs.nist.gov

    Credit

    Lawrence Livermore National Laboratory

  4. 1992

    A mirror of 36 pieces

    Jerry Nelson of Lawrence Berkeley Laboratory, with Terry Mast and George Gabor, designed the Keck telescope's 10 m mirror as 36 hexagonal segments, each 1.8 m wide and 7.5 cm thick, kept aligned by a computer; the whole was expected to weigh about as much as the Hale telescope's 5 m mirror with four times its area. Each segment needs an off-center curve that is hard to polish, so it was bent by set forces, polished as a simple sphere and released into the right shape, a method called stressed mirror polishing. The telescope took its first light on 4 December 1990 with nine segments, was completed in the spring of 1992, and holds its segments in line twice a second to 4 nm.

    www2.lbl.gov

    Credit

    Jerry Nelson, Terry Mast and George Gabor; the California Association for Research in Astronomy

  5. 2022

    Webb's mirrors beat the target

    The James Webb Space Telescope's main mirror is made of beryllium segments. Its requirement was to be diffraction limited, as sharp as its size allows, at a wavelength of 2 µm, which means a wavefront error below about 150 nm. In six months of commissioning the error measured at the main sensing point of its NIRCam camera was generally 60 to 80 nm, so the telescope is diffraction limited down to 1.1 µm, by the account of Jane Rigby and more than 600 co-authors.

    arxiv.org

    Credit

    the JWST mission and its commissioning team, reported in a paper led by Jane Rigby

Watch for

  1. Flat lenses in volume

    A metasurface, or metalens, is a flat optic patterned with structures smaller than the wavelength of light that bend light as a curved lens does; Federico Capasso's group at Harvard, with Robert C. Devlin among the authors, showed visible-light metalenses in Science on 3 June 2016. STMicroelectronics says that since 2022 it has "shipped well over 140 million metasurface optics and FlightSense modules using Metalenz IP", made in its 300 mm wafer fab in Crolles, France, and in July 2025 it signed a new license with Metalenz, the Harvard spin-out that Devlin leads as chief executive. The figure counts optics and modules together.

    Achieved whena teardown firm or another party outside the two companies confirms metasurface optics in shipping products and gives a count.

    biometricupdate.comapi.crossref.org

    Claimed by

    STMicroelectronics and Metalenz, July 2025

  2. A mirror of 798 segments

    The European Southern Observatory's Extremely Large Telescope (ELT) in Chile will have a 39 m primary mirror of 798 hexagonal segments about 1.5 m across, from 931 made in all so that segments can be swapped out for recoating, two a day. SCHOTT casts the blanks from Zerodur, a glass-ceramic that barely expands with heat; Safran Reosc in France shapes, polishes and tests them; about 9,000 edge sensors are to hold them in line to a few nanometers. ESO's timeline puts the 4.25 m secondary mirror as complete in 2026, installation of the main mirror's segments in 2028, telescope first light in 2029 and scientific first light in 2030.

    Achieved whenESO reports the full primary mirror working as one surface on the sky, with measured image quality.

    elt.eso.orgelt.eso.org

    Claimed by

    ESO, on its ELT timeline as viewed in November 2026

  3. AI agents design lenses

    Lens design is a search for curvatures, spacings and glasses that meet a specification, long done by trained designers using optimization software. Lei Sun, Kaiwei Wang, Luc Van Gool and colleagues describe LensDesigner, an AI agent that learns design rules from tasks of rising difficulty, and say it "significantly outperforms publicly available baseline algorithms" on LensArena, a benchmark of 120 design tasks that they built; they say the code will be made public. An earlier paper, OPTIAGENT (February 2026), trained a language model on lens systems from optics textbooks and called itself "the first attempt to employ LLMs in the field of optical design".

    Achieved whenthe code and benchmark are released and a group outside the authors reproduces the success rates, or a lens the agent designed is built and measured to its specification.

    arxiv.orgarxiv.org

    Claimed by

    the LensDesigner authors, September 2026

Open

  1. 01

    Graded photos of lens flaws

    An open set of photographs of real lens surfaces with cosmetic flaws, scratches and digs (small pits), each flaw marked and measured, each surface graded on the US scratch-dig scale of MIL-PRF-13830B by two trained inspectors working apart, plus a baseline model that grades from the photo and a report of how often the inspectors and the model agree.

    The brief
    Credit · unclaimed
  2. 02

    AI lens designers, rerun

    An independent rerun of published AI lens-design agents on a public task set, with every design the agents produce checked in a second, separate ray tracer and against the specification, and the results published with all files.

    The brief
    Credit · unclaimed

Open 01 · the brief

Graded photos of lens flaws

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An open set of photographs of real lens surfaces with cosmetic flaws, scratches and digs (small pits), each flaw marked and measured, each surface graded on the US scratch-dig scale of MIL-PRF-13830B by two trained inspectors working apart, plus a baseline model that grades from the photo and a report of how often the inspectors and the model agree.

Why it is still open

Under MIL-PRF-13830B an inspector matches the brightness of each scratch to calibrated reference scratches under a set lamp, and Edmund Optics notes that "because it is a subjective visual inspection the results can vary from inspector to inspector"; the more exact ISO 10110-7 method, which sizes and counts flaws, is slower and costs more. Habib Yaseen of the University of Huddersfield wrote in February 2025 that "there is no open-source optical lens dataset consisting of representative samples" for building automated inspection. The one public lens set found, for ski-goggle lenses (Dang and Wang, Axioms, 2023), labels types of flaw such as scratches and dust spots, not grades.

edmundoptics.comarxiv.orgapi.crossref.org

Built with what exists

Reject or surplus lenses from optics suppliers; a USB microscope or a Raspberry Pi High Quality Camera with a dark-field ring light; a set of scratch-dig reference standards for the human graders; Ultralytics YOLO or Meta's Segment Anything for marking flaws; images released under CC BY 4.0. Ordinary lamps light the parts; no laser is used.

  1. First step

    Buy 50 reject lenses and a set of reference standards, photograph the lenses under one fixed lighting setup, and have two people grade the same 50 without seeing each other's results.

  2. Done looks like

    At least 2,000 images of at least 300 lens surfaces; every flaw boxed and sized; every surface graded by two people; their agreement reported as Cohen's kappa; a baseline model scored against the graders on a held-out tenth of the surfaces; data, labels and code public.

Open 02 · the brief

AI lens designers, rerun

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An independent rerun of published AI lens-design agents on a public task set, with every design the agents produce checked in a second, separate ray tracer and against the specification, and the results published with all files.

Why it is still open

LensDesigner's success rates come from the authors' own runs on their own 120-task benchmark, and the paper says only that the code "will be publicly available"; on 20 November 2026 the authors' repository, created on 6 November, held one commit with a one-line README. OPTIAGENT likewise reports its own tests. No outside comparison of these agents was found in arXiv searches on 20 November 2026.

arxiv.orggithub.com

Built with what exists

Optiland, an open-source Python package for optical design and differentiable ray tracing under the MIT license; a second ray tracer, such as Ansys Zemax OpticStudio, for cross-checks; the agents' released code where available, and a general AI model such as Claude driving Optiland as a baseline (see the Note "Three working devices" on AI-designed hardware checked only by its maker); classic designs with published prescriptions, such as the Cooke triplet, as tasks with known answers. Software only; nothing is built or switched on.

  1. First step

    Write five classic tasks with known solutions into Optiland, have one general AI model attempt each three times, and check the results in the second ray tracer.

  2. Done looks like

    At least 30 public tasks, each with focal length, f-number, field of view and a target image-quality figure; every agent run at least 3 times per task; a design counts as a success only if both ray tracers agree it meets the specification within 2%; success rates, run times and all design files published.