Use case
How to design a curved waveguide for AR glasses
Yousef Arzhangnia, CTO and co-founder, Opdo · October 21, 2026 · 6 min read
Opdo's AI agents folded 21.6 mm of display optics into 7.9 mm. A curved see-through lens can be the display optic of AR glasses, if a chirped grating releases the light it guides. It is a non-sequential optics problem, designed in three iterations; both numbers are depth in front of the eye in 2D simulation, the larger one for a flat combiner.
A curved waveguide is a transparent optic, shaped like a spectacle lens, that carries display light inside its own body by total internal reflection and then releases it toward the eye.
The problem: one lens, two jobs
The challenge was to get one lens in front of the eye to do two jobs: pass the real world without focusing it, and carry a display image to the eye. Folding the display path inside the lens moves the display optics from in front of the eye to the side of the lens.
| Working spec | Value |
|---|---|
| See-through power | below 0.02 D |
| See-through distortion | below 1 arcmin |
| Shell | zero-power meniscus, R1 = 100 mm, 4 mm thick |
| Pupil | 4 mm, center moving ±5 mm |
| Display field | ±5° |
| TIR margin | at least 2° above the critical angle |
| Worst-case pupil coverage | 90% |
Shell geometry, material and eye relief are our starting values for this challenge, not fixed requirements.
The two jobs pull against each other. A meniscus with zero power for the world is almost concentric, and that same property makes it a near-perfect light trap: a guided ray keeps its angle, so it never leaves. The power difference between the two surfaces also limits how far light can travel in any curved guide, so the out-coupler has to correct it.
What we did: three iterations, three kinds of physics
Opdo's AI agents ran three iterations between 30 September and 2 October 2026. The agent picked the physics for each question, and I made the calls that needed an engineer.
First iteration: ray tracing at scale, then an explanation. The agent fixed the shell and asked how many folds fit inside. It wrote two independent 2D ray tracers that share no arithmetic, checked both against Optiland, and searched 7.5 million designs. No image light reached the eye. Instead of a bigger search, it found the reason: a guided ray keeps its incidence angle to within 1.3°, while leaving toward the eye needs a 13.3° change.
The same iteration settled the see-through rule. A zero-power shell still magnifies the world by 1.36%, so a raw deviation below 1 arcmin is out of reach for any shell of this kind. I decided to judge the residual after removing that uniform magnification, which passes to ±15° of gaze, and to continue with an out-coupler that breaks the symmetry.
Second iteration: work backward from the eye. The agent traced the wanted output ray backward to find the direction the light needs inside the glass. A flat embedded mirror could redirect the central ray, but one mirror opens only 9 to 11 mm of aperture against the 14 mm a moving pupil needs. So the agent optimized a chirped grating, and its phase-gradient slope landed within 5% of 1/R1, the curvature of the shell. Light arriving after 2, 4 or 6 folds now leaves in the same direction.
Third iteration: replace the ideal grating with physically modeled ones. The second iteration assumed an ideal grating with a fixed 15% efficiency. Here the agent switched to wave optics. It scanned 149 surface-relief structures with rigorous coupled-wave analysis (RCWA), fixed four candidates before simulating them in detail, and computed every diffraction order for both polarizations at every angle the system uses. Its own RCWA solver agreed with the open-source grcwa to 8e-4.
It also made the evaluation stricter. A dense field and pupil grid exposed holes between the 15 points the second iteration had sampled, so the worst coverage was 30%, not 52%. The agent traced the cause to the input: the prism window was 9.0 mm long, while the bounce period of the steepest field was 10.9 to 11.4 mm. The input window must be at least one bounce period long. Lengthening it lifted the 50 mm design to 85%. With PMMA, 25 mm eye relief and a 55 mm shell, it reached 89%.
Ray tracing and RCWA simulation for AR waveguides
Use ray tracing for the path through the curved body and RCWA for the grating, then feed the grating's order efficiencies back into the tracer at every hit, following each order as its own branch. This is the same multi-scale approach Ansys describes for AR waveguides. On a curved shell the local period and the incidence angle change along the surface, so the RCWA tables are indexed by both.
We ran the cheap tools first and the expensive one last. Aerospace teams call it multi-fidelity design. A fast tracer covered millions of designs, and RCWA ran on four.
Every step was checked: two tracers agree to 1e-13 mm, reversibility tests pass, and the RCWA solver passes energy conservation and reciprocity. That matters, because AI is most useful in optical design where results can be verified.
Result: 89% coverage, 7.9 mm in front of the eye
The final design reaches 89% strict and 93% calibrated worst-case pupil coverage, which meets the 90% target under the calibrated rule.
| Metric | Second-iteration design, re-evaluated | Third iteration, final | Method |
|---|---|---|---|
| Worst-case pupil coverage, strict | 30% | 89% | dense 2D evaluation, ±5 arcmin |
| Worst-case pupil coverage, calibrated | 46% | 93% | one direction per display point |
| Depth in front of the eye | 7.2 mm | 7.9 mm | flat combiner: 20.2 to 21.6 mm |
Picking the grating is a trade. A more efficient grating extracts more light at its first hits, leaves less for later ones and couples more of the world into the guide.
| Grating, corrected 50 mm design | Power to the worst pupil | World transmission |
|---|---|---|
| Ideal, 15% per hit | 2.80% of input | 91.7% |
| PG1, binary relief, 100 nm | 0.87% of input | 85.7% |
| PG3, slanted resin, 150 nm | 2.63% of input | 61.4% |
Power is the worst of 15 field and pupil samples for unpolarized light; transmission is the mean over ±15° of gaze. On the final 55 mm design PG1 keeps 87.0% of the world view.
A flat 50/50 combiner is narrower (57 to 72 mm against 80 to 85 mm), but it sits 20 to 22 mm in front of the eye and passes half the world light.
What it means for you
If you are designing AR or smart glasses and the combiner is what makes them bulky, a curved shell with a chirped out-coupler is worth a serious look. The key takeaway is this: the input window and the grating efficiency decide the result, not the fold count.
The next steps are a grating graded along the shell, a full 2D eyebox, RGB, tolerances and fabrication on a curved part. Each of these is now a defined next experiment.
Frequently asked questions
Can a waveguide for AR glasses be curved?
Yes, if the out-coupler corrects the power the curvature adds. In our design a chirped grating with a phase-gradient slope close to 1/R1 makes light from every bounce sequence leave in the same direction.
How do AI agents for optical design work on a project like this?
They pick the method per question, write and check the code, and stop at the decisions an engineer has to make. Here they used purpose-built ray tracers, Optiland, an RCWA solver and CMA-ES; on other projects they run Zemax, Code V, FRED and Tidy3D.
How long does a study like this take?
These three iterations ran between 30 September and 2 October 2026. Each one ended with a report, the code to reproduce every number, and a clear next step.
How to start
Have a similar problem? Bring us your use case. We'll deliver a production-ready design. Talk to us, or see how our AI agents work.
What would you design with AI agents?
Sources
- Draper CT, Blanche P-A. "Holographic curved waveguide combiner for HUD/AR with 1-D pupil expansion." Optics Express 30(2), January 2022. PDF
- Cakmakci O. Optical Design of Eyewear Displays. Bellingham, WA: SPIE Press, 2025. SPIE Press
- Ansys Optics. "Augmented Reality Optical System" (multi-scale simulation coupling RCWA grating models with Speos ray tracing). Ansys
- Parsonage B, Maddock C. "A multi-fidelity model management framework for multi-objective aerospace design optimisation." Frontiers in Aerospace Engineering, February 2023. PDF
- Schulz WG. "Evaluating the state of play for AI and optical design at SPIE Optics + Photonics." optics.org, 28 August 2026. optics.org
See how Opdo sources and verifies your parts.
One request in, a verified part out — routed to the right technology for your geometry, material, and tolerance.