Israel Is Building a CT Scanner for Clouds

Sometime in the coming weeks, a satellite about the size of a shoebox and roughly four kilograms, the heft of a small housecat, is set to lift off from California to attempt something no weather satellite has managed. It will begin the work of looking inside a cloud.

That sounds modest until you learn what it is in service of. The single largest source of uncertainty in how hot the planet will eventually get is not oil politics or carbon accounting. It is the cloud. The UN climate panel’s own assessment puts the likely warming from a doubling of carbon dioxide somewhere between 2.5 and 4 degrees Celsius, a range wide enough to mean either a manageable century or a brutal one, and the biggest reason it stays that wide is that we still cannot pin down what clouds will do. Clouds both cool the planet by reflecting sunlight and warm it by trapping heat, and which effect wins as the world heats up remains the single greatest open question in climate prediction.

The frustrating part is that the clouds giving us the most trouble are the small ones. The fair-weather cumulus you would have doodled as a child, a few hundred meters across, are too small and too lumpy for today’s satellites and climate models, which mostly treat the sky as smooth, flat layers. These are the small clouds “generally missed by today’s remote-sensing technologies,” in the words of Ilan Koren, the Weizmann Institute atmospheric physicist who leads the science side of the mission. We have decent pictures of giant storm systems and almost nothing on the three-dimensional guts of the ordinary clouds that drift over much of the planet.

A CT scan, but for light that bounces

This is where an Israeli-led project called CloudCT gets clever. The idea is borrowed straight from a hospital. A medical CT scanner builds a 3D image of your insides by firing X-rays through you from many angles and letting a computer reconstruct the cross-sections. CloudCT wants to do the same to a cloud: photograph one cloud simultaneously from ten different vantage points in space, then reconstruct its interior, droplet field and all.

The physics, though, is far nastier than a hospital scan, and that gap is the whole story. An X-ray travels in a straight line and simply gets dimmed on the way through. Sunlight inside a cloud does the opposite: it ricochets. A single photon bounces between water droplets many times before it escapes to a camera, so the picture you capture relates to the cloud’s structure through a tangled, nonlinear web of multiple scattering. Untangling that to recover the actual droplets is a brutal inverse problem. There is one gift hidden inside it, though: the way light’s polarization shifts when it scatters off a droplet encodes that droplet’s size, a signal our eyes cannot see but a well-built camera can.

Optical CT of clouds requires simultaneous views from space using a specialized camera. The camera is sensitive to polarization, a property of light that our eyes cannot detect, but that provides information about cloud droplets. The camera was designed specifically for CloudCT, and we will test it during the upcoming precursor mission.

That is Yoav Schechner, the Technion computational-imaging professor who handles the scanner half of the mission. The patient half, the atmospheric science, belongs to Koren at Weizmann. Their German partner, Klaus Schilling in Würzburg, has the job that sounds easiest and is not: getting ten satellites to fly in tight enough formation to aim at the same small cloud at the same instant.

Where the AI actually comes in

Reconstructing a cloud from that mess of scattered light, the old way, meant grinding through the radiative-transfer equations over and over, hours of computation per cloud. That is useless when a satellite is streaming data and the cloud you photographed has already drifted and changed shape. So Schechner’s group did what much of the frontier is doing now: they trained a neural network to do the inversion.

The result, in a system they published last year, does the reconstruction in close to real time, and it does two things worth dwelling on. First, it was trained on a vast library of physically simulated clouds, so it has effectively learned what real clouds are allowed to look like, which constrains an otherwise hopeless guess. It does not throw the physics away: the AI is taught by the physics, then runs fast in its place. Second, and this is the part I find quietly impressive, it does not just output a single number for each point in the cloud. It outputs a probability, an honest estimate of how sure it is. In a field that has been repeatedly burned by confident-looking AI that turns out to be quietly wrong, a model that hands you its own uncertainty map is exactly the right kind of engineering.

Why this is coming out of Haifa and Rehovot

It is easy, in a year of Israeli AI headlines about billion-dollar ad-measurement deals and battlefield systems, to forget that some of the most interesting work points at neither a marketplace nor a war. CloudCT is funded by a European Research Council Synergy Grant, roughly 14 million euros, one of the EU’s most competitive science awards, reserved for the rare project that genuinely needs several top labs to combine forces. Israel is two of the three: Weizmann for the clouds, the Technion for the imaging, with Würzburg supplying the satellites.

The formation flying is not a paper promise either. Back in 2021 the Technion flew Adelis-SAMSON, three nanosatellites that held formation in orbit on their own, among the first autonomous satellite clusters anyone had pulled off. The hard-won lesson there, that clever software can stand in for the size and budget you do not have, is exactly the bet CloudCT is making at a larger scale.

It also fits a quieter Israeli pattern: AI pointed at the physical world rather than at a chat box. The Israeli-founded weather company Tomorrow.io is lofting its own satellite constellation to feed AI forecasting models. CloudCT is the research-grade version of the same instinct, to build the instrument the models have been missing rather than just tune the models again.

None of this happens overnight, and it is worth being precise about what launches this summer. The shoebox going up is a single pathfinder, there to prove the camera works in space. The real scanner, the full formation of ten, is slated for 2027, and only if the pathfinder earns it. The clouds will not surrender their secrets on the first pass.

But sit with the ambition for a second. We have spent years arguing about clouds we could not actually see. A few Israeli labs decided the answer was not another argument. It was to build a CT scanner, point it at the sky, and finally look inside.