This Computer Has No Circuits. It Thinks With Hundreds of Particles in Liquid

A computer does not necessarily need transistors, silicon or even solid parts. Researchers in Germany have demonstrated a machine in which hundreds of microscopic particles, circling through a liquid and influencing one another through tiny flows, perform useful calculations.

KEY NUMBER
The experimental reservoir used 400 interacting particles, allowing hundreds of physical computing elements to operate in parallel.

A computer that uses motion instead of switches

Conventional processors control billions of transistors through carefully defined operations. The new device takes a different route. It lets a complex physical system respond naturally to incoming information, then learns how to interpret the resulting motion.

The study, published in Communications AI & Computing on July 23, 2026, was led by researchers at the Universities of Konstanz and Stuttgart. They placed microscopic silica particles in a liquid and coated one side of each particle with a thin light-absorbing carbon layer.

A focused laser warmed the coated side and propelled the particle. A feedback delay caused each particle to orbit a target point rather than move straight toward it. Arranged in a hexagonal grid, the particles disturbed the liquid around them, and those flows connected the motion of neighbouring particles.

How liquid becomes a computing network

The approach is called physical reservoir computing. Instead of training every internal connection of a neural network, engineers feed information into a naturally complex system. The system transforms that information through its own dynamics, while only a relatively simple output layer is trained.

In this experiment, the input shifted the target positions of the particles. Their velocities and locations became a rich, changing representation of the signal. The researchers then combined selected measurements to produce a prediction.

The liquid is not merely a container. Hydrodynamic flows carry information between the particles, giving the network memory and nonlinear behaviour. By adjusting particle spacing and damping, the team could tune how strongly the oscillators interacted and how long the effect of an earlier input remained.

What the particles managed to do

The system forecast two standard chaotic time series, including the Mackey-Glass signal. For a one-step prediction in one tested configuration, the normalised error was about 0.1. More established physical reservoir technologies can perform better, but this result came from a new kind of many-particle platform.

The researchers also inserted disturbances into a chaotic signal. Obvious spikes were detected with an F1 score of 0.98. More subtle anomalies, designed to preserve the signal's immediate value and slope while disrupting its hidden temporal structure, were detected with an F1 score of 0.90.

That second task is particularly interesting because many real warnings are not simple outliers. A sensor may need to notice that a pattern's history has changed even when the latest measurement looks normal.

Where this idea could matter

Physical computing may be useful near the place where data is produced. A future sensor could use the dynamics of its own material to filter a stream, identify an unusual event and transmit only the important result. That could reduce the volume of raw data sent to a remote processor.

Possible applications include machinery monitoring, physiological signals, environmental sensors and autonomous microrobots. The paper does not demonstrate any of these deployments, but it shows that interacting active matter can supply genuine parallel computation and short-term memory.

What happens next

The researchers suggest replacing elaborate laser control with active materials that respond through simpler mechanisms. Future work could also integrate sensing and computing in one small device, test real signals and measure total energy use rather than the particle layer alone.

The experiment's value is not that liquid will replace laptops. It demonstrates a broader idea: when matter already possesses memory, interaction and nonlinearity, a computer can use those properties instead of recreating all of them with software.