37 comments

[ 2.6 ms ] story [ 91.2 ms ] thread
Making CMOS sensors from scratch is very much not a thing for a thrifty startup to achieve...

Yet submitting a design to a commercial fab and saying "make it curved plz" isn't likely to work either.

So how is this done?

If you thin a silicon die enough it becomes flexible. I suspect they're using a standard planar fab then thinning the die and mounting it on some curved substrate.
They may even just be using relief cuts in the die to produce the effect. Reliefs have a number of advantages, including reducing twisting and buckling.
But not flexible in 3 dimensions...

(Ie. You can curve it around a cylinder, but not around a sphere)

Wouldn't different thicknesses allow this?
Take a sheet of paper and wrap it around a ball. Notice how it folds?

If you could take a flat surface and have it represent the surface of a sphere without warping (or vice-versa), we wouldn't have constant debates about what kind of map projections are best.

Yes, I agree. What I mean is that by varying the thickness, you can get it very close to or exactly a sphere.
(comment deleted)
Maybe you could make it like an orange peel and bend it mostly sideways? Depends on how rigid it is. Or cut ridges like a fresnel lens? Or like lasik? All depends on the material properties.
Good question. I don't know but I know from another effort that was at Sun way back in the day it is really difficult to create curvature in a system that wants flat optical masks and flat processing.

There are at least two challenges, one is curve itself, and the other is laying out the sensor to be linear across the curved surface. I'm wondering if they are going to do a Hot Chips presentation, that has been where most of the coolest advances in chip making have been detailed.

A friend speculated they might be doing a mems sort of thing where they make a 'thick' chip and then etch down to the layers of sensors, perhaps polishing, as a glass lens might be made, but for me that seems like it would be too risky. Thinning and bending is a possibility of course you need to account for the final topology in your layout.

Definitely a fun puzzle to speculate about.

Although wafers seem easy to make flat, I wonder if lithography itself might work better on a curved target.
there are lenses in the lens assembly that flattens the image. Thus if you wanted to make curved wafers you could remove these. But then the curve would be across the water itself and maybe not what you want on the individual chip level.
Interesting!

Do they start with a curved piece of silicon and do the fabbing on that? Or is it initially flat and then bent during the process?

Will this be generally isolated to very specialty wide-field imaging? And expensive, custom runs?

I would expect the sensors are manufactured flat, back thinned down to where Si is flexible and then mounted onto a curved support structure.

It looks like their US patent is here: https://patentimages.storage.googleapis.com/10/b3/ee/69d396b...

How interesting!

By the way, this brings back thoughts how astronomy used to be done with thin glass plates, doing wide area sky surveys. The plates would be placed into the plate holder, and some pressure was necessary to deform the plate to match the curvature desired.

They said you knew when you'd gone to far because you'd hear the tinkle of broken glass...

(they also baked the plates in hydrogen to increase their light sensitivity)

This is from a reply they made to a comment on the article: "Hi, you are right, we take available sensors and just change their shape before packaging. It makes it a plug-and play product, with no need to have a specific PCB development."
That’s not the first - Sony already released a product with a curved sensor in 2014, the KW1: https://www.sonyalpharumors.com/sony-confirms-kw1-perfume-se...
Right but the Sony one is a small 1/2.3-inch, this one looks like a Full Frame 35mm sensors.

Edit: I think "the first one" is the claim for "very first commercial curved sensor for a scientific application"..."for a scientific application" they said, Sony was a selfie camera :-)

So (1/(2.3)) inch == 11mm vs 35mm for the new one?
Sensor size designations are very confusing. The 1/2.3 is based on the diameter of an old tube sensor, of which a large portion couldn't be used for imaging. The 35mm is based on 35mm film, which was actually 24mm x 36mm. The only way to make sense of it is to use a table showing actual dimensions, such as https://en.wikipedia.org/wiki/Image_sensor_format#Table_of_s....
Sorry, I’m a pleb but in what capacity/setting is this particularly useful?

EDIT: I don’t mean this in a negative manner, I just don’t know about this kind of stuff so I’m curious about what settings this is useful.

(comment deleted)
As someone who have designed lenses (with Zemax et al), this means I can get by with fewer lens elements, making it less expensive. That's one aberration type I can stop worrying about.
One interesting challenge I assume they faced here is that the properties of Silicon that are relevant for transistors can be (significantly?) affected by stress/strain.

So while it seems reasonable that you could thin down a digital circuit and have it continue to function correctly when bent, a CMOS sensor is a very analog device. I wonder if noise / dark current / other properties vary depending on where the stresses are applied.

Oh, this is even more interesting after looking into it (see this diagram from another site): https://www.dpreview.com/news/7542036825/french-startup-is-p...

The curvature goes in the opposite direction of what I would have thought. I would've thought the orientation of the curve was in the sense of maintaining a "constant radius" from the end of the lens stack. (exit pupil?) I.e. the curve helps bring in the distant parts of the sensor closer and more orthogonal to the light rays.

But the diagram shows the opposite kind of curve! That makes even less intuitive sense to me.

(I can't imagine someone not involved with the company creating this diagram, or it being so wrong and not being caught by some review process? I figure the diagram is legit? Someone drawing ray tracing level details would get it so wrong?)

It depends on the lens. Some lenses will more easily want a sensor that curves away, others will prefer a sensor that curves towards them.
The image of the sensor held by the fingers looks concave.

The illustration with simplified optics shows a convex sensor.

Hard to figure it out.

I would imagine chromatic abberation and light dropoff is due to flat sensors and concave sensors would help. But I might not understand modern lens design.