Meta-Optics Replaces The Lens Stack

The whole camera is 1.68 by 1.68 by 2.11 millimetres and weighs 12.9 milligrams — about half a grain of rice. NILT's product page describes the module as comprising the sensor chip-scale package, the lens and the flex circuit, and nothing else. What it does not contain, because there is none, is a lens barrel. There is no moulded plastic doublet, no air gap, no spacer ring, no threaded housing. NIL Technology launched the metaEye on 28 October, and calls it the first camera on the market whose entire optical train is meta-optics.

That is a narrower claim than "the first metalens camera", and the narrowness is the point. A metalens is usually one flat element doing one job inside a stack that still contains conventional refractive surfaces. What NILT has put into production is a solid cube of fused optics containing two meta-optical elements, a bandpass filter and multiple apertures, bonded directly onto the image sensor's cover glass. Nothing in the optical path bends light by being curved.

The launch part is the TM28-901013, built around an OmniVision OG0TC1B — a backside-illuminated global-shutter sensor in a 1.64-millimetre-square chip-scale package, by OmniVision's own datasheet. The array is 400 by 400 pixels on a 2.2-micrometre pitch, which gives an active area of 0.88 millimetres square — the optics have to cover that and essentially nothing more, which is why the module's footprint is set by the sensor rather than by the lens. The camera operates at 850 nanometres, with the meta-optics designed for 850 ± 20 nanometres and an integrated bandpass filter rejecting everything else. Working distance is 15 to 40 millimetres.

Those last two numbers tell you what it is for. An eye-tracking camera in a pair of smart glasses sits in the rim, a few centimetres from the cornea, looking at a pupil lit by an 850-nanometre LED. It does not need to focus at infinity, it does not need colour, and it does not need to be good at anything else. It does need to be invisible, which in practice means it needs to be smaller than the hinge.

The obvious argument for flat optics is z-height, and NILT makes it: the lens cube has a total track length of 1.7 millimetres. But the more interesting argument in the launch release is not about size at all.

A conventional miniature camera is a stack of injection-moulded polymer elements held apart by air in a barrel. Polymer has a thermo-optic coefficient roughly an order of magnitude larger than glass and the wrong sign, the barrel expands, and the air gaps change with it. The focus shifts with temperature, and the shift is large enough that mobile modules compensate for it in software or with an actuator. A solid cube of bonded dielectric with the patterning done lithographically on flat surfaces has no gaps to change and a much smaller thermal coefficient to change them with. For a device that sits against a warm face in a cold street, that is not a specification-sheet nicety.

It also removes the alignment problem. Each air gap in a barrel is a tolerance that has to be held during assembly and over life. A monolithic cube bonded onto the cover glass has one alignment step — the bond — and after that the optical train cannot move relative to the sensor because there is nothing for it to move through.

The step from one meta-optical element to two is where the engineering is. A single metasurface is a thin phase screen: it can impose an arbitrary phase profile at one wavelength, which is enough to make a lens, but it gives the designer no way to trade aberrations against each other across the field. It is a singlet, and it behaves like one.

Two surfaces separated by a known thickness give you what a doublet gives you — two phase profiles and a propagation distance between them, which is enough degrees of freedom to flatten field curvature and control the chief ray angle across the image. That last one matters for a backside-illuminated sensor, whose microlens array is built with a particular chief ray angle curve in mind; a lens that does not match that curve loses light at the corners no matter how good its modulation transfer function is. NILT states that the metaEye's lens matches the sensor's chief ray angle curve, which is the kind of claim that is only available to you if you have more than one surface to work with.

The multiple apertures in the cube do the stray-light job that a barrel normally does. In a conventional module the barrel wall is the baffle; take it away and you need the baffling inside the glass.

Metasurfaces have had a long demonstration phase and a short production one, and the reason is manufacturing rather than physics. A metalens is a field of subwavelength pillars whose diameters vary across the aperture, and at 850 nanometres those pillars are a few hundred nanometres across. Writing them one device at a time, by electron beam, is a laboratory process. NILT designs and produces the optics in-house and its partners assemble the finished cameras. The company is named for nanoimprint lithography, the replication route that turns a metasurface from something written into something stamped — a master made once, then pressed out.

Eighteen months of qualification builds is the detail worth holding onto. It is roughly what a consumer optics programme takes, and it means the parts have been through reliability, yield and cosmetic screening at somebody else's incoming inspection rather than only at NILT's.

The metaEye is one configuration of a platform NILT calls metaCamera, which covers object distances from 2 millimetres to infinity in three variants: metaMacro at 2 millimetres, metaEye at 15 millimetres, and metaInfinity from 1 metre out. All three are near-infrared cameras built entirely from meta-optics, all three have been demonstrated at 850 nanometres, and the company says the design can be optimised for any near-infrared or short-wave infrared wavelength. The XY footprint is limited only by the sensor; the z-height is customisable.

That is a product architecture rather than a product, and it is the more significant announcement. A metalens tuned for one wavelength and one conjugate is a custom part. A design flow that takes a sensor, a wavelength and an object distance and returns a bonded cube is a catalogue.

For twenty years the argument against metasurfaces in volume optics has been a list: narrow bandwidth, poor off-axis performance, no manufacturing route, no way to correct aberrations with a single phase screen. The bandwidth objection is answered here by not needing bandwidth — an 850-nanometre illuminated scene with a 40-nanometre design window is exactly the case where a diffractive solution is allowed to win. The aberration objection is answered by using two elements. The manufacturing objection is answered, if the ramp holds, by nanoimprint.

What is left is a question for everyone who polishes, moulds or coats small optics for near-infrared sensing. The metaEye does not beat a conventional lens on image quality, and it is not trying to. It wins on volume, height, mass, thermal stability and the absence of an assembly step — and it wins in an application where the required image quality is low enough that those things decide. That application is now a consumer one, ramping into production for headsets and glasses, and the list of other applications that look like it is not short.