In a pot of water, microscopic spheres, colloidal particles, are floating. Each sphere is coated with long polymer chains that look like a brush and carries an electric charge. If you get the temperature and solution composition right, the spheres start arranging themselves into ordered patterns — crystals. But here's the catch: under the same conditions, you can get either a dense packing, like oranges in a crate, or a loose one, like bubbles in foam. Standard ways of telling one pattern from another fall short when the solution contains a mix of everything at once and some spheres clump into shapeless blobs.

The authors came up with a way to look at the environment around each particle and compare how similar it is across different spheres. Previously, researchers used local environment descriptors: they distinguish packing variants well, but they're too slow for computer simulations. So the scientists turned to neural networks: an SE(3)-equivariant transformer (a model that doesn't confuse rotated or shifted structures) was trained to quickly predict how similar one sphere's environment is to another's. It's like instead of manually measuring distances between all the oranges in a crate, you train a robot to instantly tell whether the arrangement matches a sample.

Using this neural network, the researchers were able to compute the free energy (a measure of stability) for different crystal packings. It turned out that by changing the surface charge of positive and negative particles while keeping the same attraction between them, you can switch the system between two types of crystals. This is like tuning a radio: turn one knob and you pick up one station, turn another and you hear completely different music, even though the radio itself is the same.

Now, knowing the energy landscape, you can calculate in advance what conditions are needed to get exactly the structure required for a specific application. Instead of a long trial-and-error search for recipes, the experimenter gets a map: here a type A crystal will grow, and here type B. The work opens the way to rational design of colloidal materials, where properties are set not by chemical composition but by how particles are packed relative to each other.