A Powder That Remembers Motion express

Original: "Learning Associations in Reconfigurable Particle Packings via Local Cyclic Driving"
arXiv:2603.14534 · 2026-03-15 · CC BY 4.0 · 1 min
Granular particles with no brain or wires learn associations — like Pavlov's dog, but made of sand.

A handful of sand on a table is just a pile. But if you shake the grains just right, the pile starts behaving like a memory: it remembers which motion follows which. In the new experiment, particles were moved cyclically, and after training, they reproduced the desired pattern even without a full command.

The secret lies in the contacts between particles. When grains shift, they change neighbors and interaction forces — so the pile's 'wiring' rearranges itself. This is like how a path in the woods becomes more visible if you walk it often: each step changes the ground, and the route is remembered.

The most surprising thing: the system has not a single wire or sensor. All learning is just particles rearranging under external motion.

The scientists tested tasks of varying difficulty. The simple task was solved right away — the particles were already moving in the right direction. In the hard task, learning barely helped until pauses were added: the particles had time to 'settle' and remembered better. And in the medium task, simply changing the arrangement of inputs and outputs did the trick.

This approach is like training muscles: without rest, muscles get overworked and don't grow, but with breaks, they strengthen.

The experiment shows that associative memory doesn't need complex circuits — just a granular medium and the right rhythm. Perhaps in the future, we'll see materials that learn on the fly, like this sand.

In the experiment, particles moved without heating — only through mechanical action, as if you were shaking a box of marbles.

Scientists
Andrey Markov
Concepts
Reinforcement Learning learning neural network memory architecture protein language model self assembly reservoir computing machine learned interatomic potential autoencoder entropy driven self assembly training atomistic simulation parameter memory effect neural network data reconstruction quantum entanglement training agent one dimensional chain brain rhythmic transport information storage attention mechanism