2026-09-20
Nature Reviews Physics surveys SOMNs that rewrite conductance in-field and learn via reservoir computing and associative feedback.
Running neural-network software on GPUs is already an energy problem. Memristor crossbars can fold matrix-vector multiplies into hardware, yet each cell still has to be addressed and the weights still arrive from a global training loop. Brains do not work that way. Wiring grows, synapses change, and there is no weight file to download.
This Nature Reviews Physics Perspective names a class of bottom-up nano-circuits self-organising memristive networks (SOMNs). Silver nanowires, or tin, gold and silver nanoparticles, land on a substrate as a dense random mesh. Junctions are resistors with memory. An electric field grows metallic filaments or atomic hillocks, and conductance follows. Authors from Los Alamos, the University of Pisa, Italy's metrology institute, UCLA, Politecnico di Torino, the University of Canterbury and the University of Sydney argue that learning can live in the material.
Two material routes, different microscopics, similar macroscopics.
Silver nanowire networks switch by electrochemical metallisation. An applied field drives Ag+ across a polymer or chalcogenide coating; the ions nucleate as metallic filaments and the junction conductance drops in integer multiples of the quantum \(G0=2e^2/h\). Remove the voltage and interfacial energy plus a nanobattery effect dissolve the filament. Short-term memory sits in the millisecond-to-second window. Stronger bias can rupture a wire by Joule heating and then electrochemically rewire the gap, which is structural long-term memory. The review calls the first process reweighting (an existing junction changes conductance) and the second rewiring (a new junction appears across a gap).
Nanoparticle networks are usually deposited near the percolation threshold. In the best-studied tin systems the driver is field-driven atomic rearrangement, not electrochemistry: low fields grow surface hillocks that modulate tunnelling; high fields close the gap into a filament. Timescales are microseconds. One gap switching redistributes voltage and current, which can trigger avalanches.
Individual junctions need not be wired out. A multi-terminal electrode set reads a conductance matrix, each entry the effective resistance between a pair of pads. Kirchhoff's current and voltage laws couple local updates into global feedback. Theory writes each junction as a first-order memory ODE \(dg/dt=F(g,\delta v,t)\), then folds the circuit constraints in with projector operators \(\OmegaB\) and \(\OmegaA\). Coarse-graining yields motion of the mean conductance \(\langle g\rangle\) in an effective potential \(V{\Delta v}\): raising voltage can split one well into two, a low-conductance phase and a high-conductance phase. Mean-field tracks experiment in the subcritical regime. Near the transition the Lyapunov exponent turns positive and mean-field fails.
Learning does not backpropagate into every weight. Two schemes already exist in hardware.
The slogan, borrowed from a 2024 Neuron essay, is that the hardware is the software.
This is a Perspective, not a new leaderboard. Evidence is assembled from published experiments and simulations.
On the physics side, conductance avalanches in nanoparticle networks are self-similar across timescales. One trace covers 2400 s; successive zooms stretch time by 10, 100 and 1000 and the vertical scale by 3, 9 and 27, and the shape still looks the same. The probability density of \(\Delta G\) is heavy-tailed, inter-event intervals follow a power law, and the autocorrelation of switching decays as a power law over several decades. Shuffle the interval sequence and the decay steepens. The same power laws survive a 1000-fold faster sampling rate and even microsecond sampling. Nanowire networks sit above percolation density and show avalanche criticality just above a critical voltage; well above that voltage they enter a discontinuous "bolt" phase akin to dielectric breakdown. Measured exponents do not match known universality classes.
On plasticity, stimulating one electrode pair potentiates that pathway and then relaxes; temporally correlated pulses produce paired-pulse facilitation; unstimulated pathways still change, a heterosynaptic effect from Kirchhoff redistribution. Nanowire networks can freeze short traces into longer structural rewiring; tomography sees engram-like conductance paths. Nanoparticle plasticity lives on microsecond scales and, the authors say, still needs work.
| Quantity | Nanowire networks | Nanoparticle networks |
| Short-term memory | milliseconds to seconds | microseconds |
| Criticality | just above a critical voltage | fabricated at percolation |
| Microscopic driver | electrochemical filaments | field-driven atoms and tunnelling |
On tasks, physical reservoir computing has been shown experimentally on waveform regression, MNIST classification, spoken digits, time-series prediction, speaker recognition and Braille classification. Figure 7 encodes MNIST digits 0, 3, 5 and 9 (100 training images averaged per class) as voltage pulses; two readout channels separate the digits, while the conductance channel carries internal memory and is more useful for sequence memory than for classification. The Perspective does not report accuracy numbers for these tasks. Simulations find that a limited electrode set already approaches the upper bound of reading every node. On the associative side, an n-back working-memory task can consolidate short-term traces into longer-term memory with positive and negative feedback, in a manner close to contrastive learning.
Mean-field short-term plasticity curves match nanowire pulse experiments and grid-graph simulations. Near criticality, simulation and mean-field diverge.
For people building edge models, the pitch is not another benchmark point. Computation and memory sit in the same disordered nanomaterial, so there is no per-device addressing and no global backpropagation. Crossbars remain the right tool for running a trained network. SOMNs are aimed at settings where the environment keeps changing, bandwidth and power are tight, and the model has to retune on the device. Robotics, continuous sensing and personalised healthcare are named as targets.
It is a map of a research direction, not a chip datasheet that replaces a GPU. Reservoir computing still often trains the readout in software. A scheme that truly needs no external training is an open question. Mean-field fails in the critical regime that is supposed to be computationally richest. No joule-level comparison with digital accelerators is given.
The authors list the gaps themselves. Energy use on real tasks is unmeasured. More self-contained learning, without an external readout, is missing. How many states these architectures can store is unknown. Mean-field collapses at the transition, so tools such as self-organised criticality are needed. Critical exponents sit outside standard percolation classes. Nanoparticle plasticity is under-studied. Kirchhoff dynamics and electrode-placement rules for true multi-terminal operation are incomplete. Whether equilibrium propagation and related physical-learning rules can be implemented reliably in hardware, including away from the quasistatic limit, is still open.
A few claims sit on thin ice even inside the review. The energy-crisis framing is the motive, yet the text never quotes joules, watt-hours or a digital baseline; the efficiency argument is physical intuition, not a measured bill. MNIST and related tasks are shown as feature separation, with no accuracy table against a digital reservoir or a crossbar of similar size. Nanowire and nanoparticle junctions differ by an order of magnitude in time, yet they are told as one learning story; transfer across materials is not closed here. Corresponding author Caravelli now works at Planckian. The Perspective is an academic survey; the path to a deployable device still runs through lab devices and theory.