To run today's massive language models, data centers consume as much electricity as an entire city. We have long believed that computation must be performed by precisely controlling the movement of electrons on electronic circuits built from meticulously arranged logic gates. Yet this very obsession with precise control has become a trigger for the current energy crisis. In response, efforts to repurpose the complex physical phenomena of the natural world as computational resources have quietly been expanding their reach.
The Obsession with Precise Control Wastes Energy
The Limits of the von Neumann Architecture
The von Neumann computer, devised in the 1940s, forms the foundation of virtually every modern information device. This architecture separates the processing unit from the memory unit, advancing computation by shuttling data back and forth between the two. The time and energy wasted in this data movement is known as the von Neumann bottleneck, and it has long stood as a major barrier to improving computational speed. Modern artificial neural networks, with parameter counts reaching tens of billions to trillions, must rigorously compute and continuously optimize every single connection weight. The very philosophy of trying to perfectly manage the behavior of every electron on a transistor is now colliding with the limits of hardware and power.
The Rise of Reservoir Computing, Which Harnesses Physical Phenomena
As an approach to break through this impasse, a computational method called reservoir computing has been studied since the early 2000s. Whereas ordinary neural networks learn every connection from input to output, this method fixes the complex network structure of the intermediate layer—the reservoir—and uses it as is. Input data undergoes complex nonlinear transformations inside the reservoir, generating a rich variety of states. All that remains for humans to do is tune the output layer, which picks out only the necessary information from among these complex states. Because there is no need to compute all the internal weights, the computational cost of training drops dramatically.
This theory later evolved into physical reservoir computing, which directly harnesses the complex dynamics of the natural world for computation—from ripples on water surfaces to interference patterns of light and spin waves in magnetic materials. Early experiments, such as filming the surface fluctuations of a bucket of water with a camera and recognizing input audio data from the shape of the waves, demonstrated that physical phenomena themselves possess information-processing capability.
The Dilemma Posed by the Brute-Force Approach of Time Multiplexing
However, existing physical implementations carried a particular dilemma. Because the physical interactions inside the system cannot be freely altered from the outside, it is difficult to secure a sufficient number of variables (computational dimensions). With too few variables, complex problems cannot be solved. To compensate for this, a technique called time multiplexing has been used, in which a single input data stream is finely divided and fed into a single physical node multiple times with time delays.
This technique of artificially increasing the number of variables over time does indeed improve computational power. However, it cancels out the inherent advantage of parallel processing—namely, that the entire system responds the instant data arrives. The sequential procedure of feeding data in one piece at a time and waiting for each response before moving to the next has become a serious bottleneck on processing speed in modern applications that demand fast, real-time processing.
Fully Parallel Computation Woven by Liquid and 400 Particles
A Fluid Network Formed by Active Colloidal Oscillators
Research teams at the University of Konstanz and the University of Stuttgart broke through this constraint not by using silicon or optical elements, but by harnessing the collective motion of microscopic particles. The stage they chose for computation was a microscopic space, only tens of micrometers across, filled with liquid.
The research team arranged approximately 400 active colloidal oscillators in a cell filled with liquid. A colloid refers to a microscopic substance dispersed in a liquid, such as fat globules in milk or carbon particles in ink. When given external energy, such as light or an electric field, these particles respond spontaneously and begin periodic oscillation.
The central mechanism at work here is a phenomenon called hydrodynamic coupling. When one particle moves, it pushes the surrounding viscous liquid, and the resulting flow influences the motion of neighboring particles. Even without direct collisions between particles, the movements of all 400 particles become instantaneously coupled through the continuous medium of the liquid.
Converting the Interference of Ripples into a Computational Resource
When multiple stones are thrown into water in succession, countless ripples spread out and collide, producing complex interference patterns. By precisely observing the intricate shape of these ripples, one can infer, in reverse, when and how large a stone was thrown in. The research team harnessed the flow of the liquid and the collective motion of the particles as exactly this kind of network of ripples.
When external input data is thrown in like a stone, the entire particle population responds through the fluid medium, naturally transforming the characteristics of the original data into a complex, high-dimensional pattern. Variables such as the position and velocity of the 400 particles become intertwined through fluid interactions. This interaction grows exponentially more complex as the number of particles increases, providing the vast state space required for advanced information processing.
Predicting the Future of Chaos with a Single Readout
The Rules of Ripples Extracted by a Linear Filter
How does this fluid-based physical system perform computation? Its essence can be condensed into the following simple relationship.
Here, denotes the weighting matrix in the output layer. This equation shows that by multiplying the complex overall motion of the system, , by a linear filter that determines which features to focus on, one obtains the desired answer, .
The positions and velocities of roughly 400 particles vibrating in the liquid form the vast state space . The arrangement of the particles changes moment by moment in response to the input data, and the history of that change influences the current state. All that humans need to do is a simple additive adjustment at the output stage—deciding which parts of this complex pattern of ripples to weight most heavily.
A 50-Fold Speedup and High-Precision Prediction Compared to Conventional Methods
The most significant achievement is that the time multiplexing required by conventional implementations has become completely unnecessary. Rather than reusing a single node over and over, all 400 particles operate simultaneously as computational resources. A single readout operation is sufficient to complete a full parallel computation. As a result, data-processing latency, which typically runs around 50 milliseconds in conventional physical reservoirs, was reduced to under 1 millisecond—achieving roughly a 50-fold speedup.
Using this fluid reservoir, the research team performed predictions on chaotic time-series data governed by the Mackey-Glass equation. This is an extremely difficult task for intuitive prediction, since past states exert complex influences on future ones. Without using any time multiplexing, and relying solely on a single readout from the system, the team calculated values 50 steps into the future. Compared to a conventional single-node reservoir based on time multiplexing, the normalized root mean square error—an indicator of prediction deviation—was reduced from 0.15 to 0.02, a decrease of roughly 87 percent. This serves as evidence that the fluid's motion flawlessly handles both the memory of past data and the required nonlinear transformations.
| Comparison Axis | Conventional Neural Network | Existing Physical Reservoir | Fluid Colloidal Reservoir (This Study) |
|---|---|---|---|
| Computational Substrate | Electronic circuits on silicon | Single node, such as optical elements or magnetic materials | Roughly 400 microscopic particles vibrating in liquid |
| Internal Control | All connection weights rigorously computed and optimized | Interior treated as a black box | Interior treated as a black box |
| Parallelism of Processing | Dependent on architecture and hardware | Pseudo-high-dimensionalization via time multiplexing | Fully parallel processing, not reliant on time multiplexing |
| Adjustment of Interactions | Parameter changes in software | Requires redesign of physical structure | Coupling strength and memory retention time adjustable in situ |
| Computational Speed (Latency) | Dependent on architecture | Typically around 50 milliseconds | Under 1 millisecond (roughly 50-fold speedup) |
Capturing Subtle Anomalies That Slip Through the Net of Statistics
Detecting the Breakdown of Hidden Temporal Correlations
An equally important target of verification was the detection of hidden anomalies buried within data. Data collected in the real world always contains noise. The researchers introduced a special type of anomaly signal into the input data—one in which instantaneous statistical properties such as the mean and variance remain completely unchanged, while only the temporal correlations subtly break down.
Conventional rule-based monitoring systems judge anomalies based on whether a certain threshold has been exceeded. If the overall mean and variability remain unchanged, such systems mistakenly conclude that operation is normal. However, a slight disturbance in the order or rhythm in which data fluctuates is often a precursor to major structural failure or system breakdown.
The collective dynamics of particles mediated through the fluid have the property of determining the next motion while carrying over traces of past motion. Consequently, when the rhythm of the input signal is even slightly disrupted, an irreparable deviation arises in the particles' oscillation pattern. Leveraging this characteristic, the research team classified the occurrence of anomalies that involved no change in mean value with over 99 percent accuracy, precisely extracting subtle underlying pattern changes.
Dynamic Adjustment of Memory Retention Capacity via External Driving
What underlies this high level of accuracy is the flexibility afforded by control from outside the system. By adjusting the external driving force, it became possible to change, on the fly, both the coupling strength between particles and the duration for which past inputs are retained as memory within the flow of the liquid.
To perform complex computation, there is no need to fully elucidate all the physical laws governing the interior of the system, nor to rigorously control the trajectory of each individual particle as one would control a transistor. As long as the system as a whole returns a reproducible and reliable response to input, that physical phenomenon can be directly harnessed as a computational resource.
Can the Fluctuations of the Laboratory Be Carried Out to Real-World Sensors?
Prospects Toward the Ultimate in Energy Efficiency
The components of a computer need not necessarily be silicon. By entrusting information processing to a complex physical system itself, it now appears possible to suppress the power consumption of training and inference to the microwatt scale—far below that of conventional electronic computing units—potentially raising overall system energy efficiency by more than a thousandfold.
This points toward the development of edge devices in which the sensor itself performs sophisticated information processing right where the data is generated, rather than sending data to a distant, massive data center for processing. It connects directly to applications such as real-time detection of early-warning signals for major earthquakes and weather disasters, extracted from noisy real-world observational data—whether the subtle readings of seismometers or the vast records of climate change.
Verifying Room-Temperature Operation and Resilience to External Noise
At the same time, clear challenges remain on the path toward practical application. The current system remains at the stage of a model case within a tightly controlled laboratory environment. In real-world settings, unnecessary noise—such as abrupt temperature changes or unexpected external physical shocks—would inevitably be added to the system. It remains unknown just how stably a computer built upon the premise of liquid fluctuations can continue to operate in the face of such unwanted physical disturbances.
There is also the hurdle of integrating microchannels on the scale of tens of nanometers onto a chip, while miniaturizing the device without sacrificing its energy-efficiency advantages. This fluid computer, which poses the fundamental question of harnessing the complexity of the natural world directly for computation rather than suppressing it, awaits its next stage of verification—the day it moves beyond the confines of the laboratory to operate in real-world environments.