Lens Design Using Artificial Intelligence

In what should be a surprise to nobody, companies that make camera lenses have started turning to Artificial Intelligence for their designs. Have you ever wondered how modern lenses are getting so much sharper and lighter than even a few years ago?

Figuring out how to combine pieces of glass to create a lens of a particular focal length (or a zoom) is unbelievably complicated. Until computers were available to help lens designers, camera lenses were really, really, bad.

It used to take teams of lens designers many years to come up with a viable lens. They would have to begin by imagining the number, shape, and composition of the lens elements, and then do light ray-tracing calculations to find out if a decent image would get rendered onto the film or sensor.

It was many decades before anybody even attempted a zoom lens design. It was another few decades before serious photographers would even consider buying a zoom lens. Nowadays, some zooms are within a whisker of being just as good as a fixed focal length lens.

Combine lenses to rid chromatic aberrations

(Image courtesy of phys.libretexts.org)

A typical equation used in light ray-tracing

(Courtesy of phys.libretexts.org)

Try to imagine performing calculations like what’s shown above millions of times over as you adjust multiple lens element shapes, spacing, and glass materials with different refractive indices. It’s miraculous that any decent lenses exist at all.

Next, consider focus. The light rays shown above no longer come into the lens in parallel, because the subject isn’t at infinity. For near objects, the light rays come in more like a cone, with the peak of the light cone at the subject. The lenses now require some moving elements to focus things that aren’t at infinity.

A near subject: more complicated

(Courtesy of http://hyperphysics.phy-astr.gsu.edu/)

Notice the “lensmaker’s equation” above has some “R” terms, that assume the lens shapes are slices from a perfect sphere with a well-defined radius. Modern lenses usually include aspherical shapes, with complicated functions replacing the simple “R” terms. Now, a designer’s job just got a whole lot tougher.

Fresnel lens (courtesy EdmundOptics.com)

I’m predicting that in the future, companies will come out with lenses comprised mostly (entirely?) of fresnel lenses, including negative fresnel lenses (with concentric rings of troughs in glass instead of raised rings in the glass) and even the equivalent of aspherical fresnel lenses, where the concentric rings aren’t all the same height. This type of lens could be extremely light and well-corrected. Talk is cheap, however, since I have no idea how difficult it would be to manufacture the precision ‘troughs’ in glass.

Companies and research efforts are increasingly incorporating AI (including machine learning and deep learning techniques) into ray tracing for camera lens design, simulation, and optimization. This isn't yet ubiquitous in every major lens manufacturer like Canon, Nikon, or Sony for their consumer camera lenses (based on public info), but it's an active and growing area in optical design software, specialized firms, and academic/industry collaborations.

Ray tracing is the standard method for simulating how light rays propagate through lenses to evaluate aberrations, image quality, etc. Traditional ray tracing in tools like Zemax (now Ansys OpticStudio), CODE V (Synopsys), or LightTools is computationally intensive, especially for complex systems or high-volume optimizations. AI helps by:

• Accelerating simulations (e.g., via differentiable ray tracing, where gradients enable faster optimization).

• Automating lens design (inverse design: specify desired performance, and AI proposes lens configurations).

• Enabling end-to-end optimization that combines optics with computational imaging (e.g., pairing lenses with AI post- processing).

Here are some key examples of companies and approaches involved:

Paraxial Optics offers an AI-powered optical design platform that uses differentiable ray tracing and hybrid AI tolerancing, claiming 10–100× faster workflows for optical engineers designing lenses and systems.

Peak Nano developed HawkAI, a prompt-driven AI tool that leverages machine learning to test millions of lens permutations, configurations, and materials for optimized prescriptions—aimed at revolutionizing optics design while integrating with traditional tools.

3DOptix provides an optics simulation platform with GPU ray tracing and an "Optics AI search copilot" for design assistance.

Anax Optics specializes in automated optical design using inverse ray tracing, topological optimization, and AI.

Larger players like Ansys (Zemax OpticStudio) and Synopsys (CODE V, etc.) support advanced ray tracing for lens design, and the field is evolving toward AI integration (e.g., for optimization and multiphysics simulations), though not always strictly branded as "AI ray tracing."

Research institutions have produced methods like DeepLens, which uses deep learning and differentiable ray tracing to autonomously design lenses (including computational ones with extended depth-of-field) from flat surfaces—highlighting AI's potential to transform refractive optics design.

Tools like NVIDIA's OptiX enable GPU-accelerated ray tracing used in scientific optical modeling (including camera/lens sims). GPU hardware will immensely improve the speed of modeling, to allow vastly more what-if efforts.

Artificial intelligence is a hot frontier in optics—AI doesn't fully replace expert designers yet, but it's making ray-tracing-based lens work faster, more automated, and accessible.

The camera companies that most fully adopt artificial intelligence are going to be the winners, while companies that ignore this approach or wait too long to adopt it are doomed to wither and die.