Cloud cameras see flat snowflakes. AI is trying to work out their invisible third dimension

When an aircraft travels through an ice cloud, a scientific camera can record the silhouettes of individual crystals. Some look like plates; others have branching arms or complicated bullet-rosette shapes. But an image is flat. Two particles whose shadows look similar can have quite different masses and surface areas, and those hidden differences help determine how clouds scatter sunlight, absorb infrared radiation and produce precipitation. Meteorologists have long had to infer difficult three-dimensional properties from limited measurements.[1][2]

Researchers led by J. Ko at Columbia University and collaborators have built a machine-learning approach to that missing dimension. The original paper appeared in the Journal of Geophysical Research: Machine Learning and Computation in December 2025, and the US Department of Energy highlighted it on 20 October 2026. The work addresses a measurement bottleneck rather than introducing a weather-prediction chatbot: can an algorithm estimate the mass and morphology of a tiny ice particle from the view that a cloud camera is actually able to capture?[1][2]

Teaching a model with crystals that have known bodies

The team began by constructing synthetic three-dimensional ice crystals with geometric characteristics informed by material from the Department of Energy's 2021 Ice Cryo-Encapsulation Balloon, or ICEBall, campaign. Because the computer generated those particles, researchers knew the properties that would be invisible to a real two-dimensional instrument. They then rendered particle images and trained neural networks to predict attributes including effective density, surface area and the number of branches or 'bullets' in rosette structures.[1][2]

This is a clever answer to a familiar supervised-learning problem. When you photograph a real tiny ice particle, there may be no easy way to measure its complete shape at the same instant. The missing labels make it difficult to train an algorithm directly. Simulated crystals offer paired examples: a known three-dimensional object and its projected image. The critical test is whether the images are sufficiently realistic for lessons learned from them to transfer to natural cloud particles.[1][2]

The paper reports high predictive skill on previously unseen synthetic images. For single-view tasks, a ResNet-18 architecture achieved R-squared values around 0.99 and 0.98 for two microphysical properties, and mean absolute error of 0.10 on the number of crystal bullets. An R-squared of 0.99 means the fitted predictions explained about 99% of the variance of that target in the examined simulated test distribution. It does not mean a 99% probability of correctly weighing an arbitrary real cloud crystal. Nor does it describe how frequently a weather forecast will be right.[1][2]

Why tiny shapes influence the planet

Clouds are central to Earth's energy balance. An ice particle's geometry affects how rapidly it falls, how it grows and what directions it scatters radiation. A plate-like crystal and a complicated rosette may have similar maximum dimensions while presenting different surface areas and masses. Cloud models therefore use simplifying relationships between particle size and other properties, some of which are uncertain because the needed measurements are sparse.[1][2]

If an algorithm could extract more reliable microphysics from airborne imagery, researchers could refine those relationships and compare the effect on modelled cloud behaviour. This is an indirect but potentially important path from computer vision to understanding climate. It does not mean a new neural network has already improved tomorrow's forecast, reduced climate uncertainty by a quantified percentage or reconstructed the actual hidden volume of every photographed ice crystal.[1][2]

The important missing experiment

The next step is to test accuracy against genuinely measured particles from different cloud conditions, instruments and growth regimes. Synthetic rosettes may omit broken crystals, irregular branching, optical artefacts or mixtures not anticipated by the generator. Even a striking simulated result can degrade when confronted with unfamiliar reality. A second camera angle, when available, may make some hidden structures easier to infer; the paper reports stereo-view improvements on its benchmark, but practical access to such observations is a separate engineering question.[1][2]

The scientific achievement is to turn an otherwise unlabelled measurement problem into a controlled learning experiment. Whether that synthetic training ground captures the complexity of the atmosphere is the decisive next test. Sometimes the most consequential AI innovation does not forecast the weather directly. It helps scientists measure the small objects from which their weather models are built.[1][2]

Report an error or suggest a correction