AINEWS
A 400 particle reservoir computer (a physical swarm whose motion does the math) hit 0.90 F1 (a standard accuracy score) on a hard anomaly task.
A drop of water-lutidine mixture, heated to 28°C and steered by a green laser, just did a job usually reserved for memristor chips: it forecast a chaotic Mackey-Glass series and flagged anomalies that leave a signal's mean, variance, and short-time autocorrelation untouched. The catch, which the Konstanz and Stuttgart team put on the first page of their paper, is that the 400-particle reservoir is roughly ten times less accurate than the memristor systems it implicitly competes with. The Tom's Hardware re-report makes the comparison explicit.
A reservoir computer is a physical system whose dynamics do the computation; the designer trains only the readout, not the substrate. That is why the same forecasting and anomaly-detection benchmarks recur across the field: a photonic crystal, a memristor crossbar, or 400 silica spheres orbiting in a water drop can be scored on the same yardstick.
The Konstanz-Stuttgart substrate is a swarm of 400 silica spheres capped with 80 nanometers of carbon. A 532-nanometer laser, steered by a two-axis acousto-optical deflector scanning at 100 kilohertz, drags the particles into patterned orbits. Real-time microscopy tracks each sphere, and the "computation" is the way the spheres' motion, coupled through the surrounding fluid, responds to an input signal. A companion paper (arXiv:2601.05767) establishes the underlying objects as tunable colloidal swarmalators held together by hydrodynamic coupling.
On the canonical Mackey-Glass one-step forecast, the array reached a normalized root-mean-squared error of about 0.1. On the harder anomaly-detection task, where the input has the same mean, variance, and short-time autocorrelation as the training distribution, it scored an F1 of 0.90. The paper states the reservoir "does not outperform established physical implementations," and the Tom's Hardware re-report makes the gap explicit: time-multiplexed memristor reservoirs reach NRMSE of 0.01 or better on the same benchmark.
Time-multiplexing, in a memristor reservoir, is the trick of feeding a single physical node a sequence of delayed input values to create a large effective network from one device; it is the reason memristor reservoirs look like a thousand oscillators when they are really one. A decade of concentrated optimization sits behind the memristor numbers. The hydrodynamic array is on the other side of that trade: ten times less accurate, and a substrate that does not need time-multiplexing to expose its dynamics.
Three tunability axes carry the comparison. Lattice spacing sets how strongly the particles couple through the fluid, a damping threshold sets how freely the spheres swing, and the input can reach as few as 20 percent of the oscillators without breaking the benchmark. Forecasting error varies by more than a factor of three across that parameter space, which means a designer can re-tune the substrate at runtime. Nanowerk's coverage flags the same point: the hydrodynamic array is a substrate that exposes more knobs, not a faster chip.
Accuracy held when individual particles stopped responding or clumped together, failures a memristor crossbar would register as a fabrication defect. The 400-particle array absorbs the loss and keeps running the benchmark; a memristor crossbar cannot, because its result is built on a single time-multiplexed channel that has to stay intact.
The January arXiv preprint framed the platform as positioned "apart from nearly all existing physical reservoirs" by avoiding time-multiplexing. The peer-reviewed paper is more candid, and more useful. The field has gained a new data point: another physical substrate, another set of tunable parameters, and a self-critique written on the first page. The next watch item is whether the swarmalator platform can close any of the ten-times gap without giving up the runtime tunability and fault tolerance that justified building it in the first place.