Weather, ocean currents and other chaotic systems are difficult to predict for a simple reason: small changes can grow quickly, while events in one region can influence what happens far away. Reservoir computing is well suited to such time-varying signals, but most physical reservoirs are built with fixed connections. Once fabricated, their hardware cannot easily adapt when the dynamics themselves change.
Why fixed networks struggle with a changing world
To address this mismatch, the research team developed a reservoir-computing system whose connectivity can be physically reconfigured during operation. The platform is based on a 16-kb self-rectifying memristor array and a small-world network architecture, which combines strongly connected local regions with a small number of long-range links. This allows the system to focus on locally active regions while still exchanging information across distant parts of the network - a useful combination for multiscale chaotic dynamics.
One memristor, three functions
The key is that the same memristor can play different roles depending on the applied voltage. At low voltage, it provides nonlinear analogue responses for local processing. At higher voltage, stable resistive switching turns selected devices into routing units that create long-range connections. Under reverse bias, self-rectification suppresses unwanted current so inactive devices can remain in a low-leakage standby state. In device-level measurements, the memristors showed a rectification ratio above one million and up to 128 distinguishable conductance states. Rather than asking software to imitate a changing network on static hardware, the physical array itself can change how its nodes are connected.
From the butterfly effect to forecasting chaos
The team first tested the platform on the Lorenz system, the classic mathematical model associated with the "butterfly effect". The reconfigurable reservoir achieved a normalized mean squared error (NMSE) of 0.0613 with a core-array energy cost of 0.83 pJ per selected operation. It was then applied to 30 days of atmospheric cloud-thickness dynamics. By concentrating computing resources on strongly changing regions and dynamically linking them, the system maintained more than 90% frame-by-frame prediction accuracy and improved accuracy by 7.6% over fixed-topology reservoir hardware. The results suggest that network topology can become a physical, adjustable property of computing devices themselves, offering a route towards energy-efficient neuromorphic hardware for forecasting other nonlinear, multiscale systems.
National Science Review
Experimental study