A selector shrinks to five nanometres, and the recording array still will not use it

What the work claims

The paper reports that a selector device based on the negative-differential-resistance effect in chromium-doped vanadium sesquioxide keeps its switching behaviour when the active film is thinned all the way to five nanometres, and that the thin device is in some respects better: lower leakage current and a more abrupt transition than thicker films.¹² Concretely, the five-nanometre device holds its leakage near one hundred nanoamps below half a volt, then snaps on across a window of order ten millivolts near a threshold of about one to one and a half volts, after a one-time forming step near two volts.¹ A second, more mechanistic claim rides underneath. At these thicknesses crystalline and amorphous films start to behave alike, and both require a one-time forming step before they switch. Transmission electron microscopy pins the likely cause on a thin amorphous layer that grows at the interface with the titanium-nitride electrode, and elemental mapping shows both a complex distribution of the chromium dopant and diffusion of titanium into the film from the electrode.¹

This is a device and materials characterisation: electrical measurements plus electron microscopy on fabricated films. It is a primary result, narrow and concrete, and it makes no system-level or biological claim at all. Everything I say below about microelectrode arrays is my extrapolation from the device physics, not a claim the authors make.

How it works

A selector solves a wiring problem. If you want to address many devices with few wires, the cheapest layout is a passive crossbar: a grid of row and column lines with a device at each crossing, and no transistor anywhere. The flaw is that current does not politely take only the path you selected. It leaks through the half-selected devices that share a row or a column, these sneak paths sum across the whole array, and past a certain size the leakage swamps the one cell you meant to read. A selector is a two-terminal device placed in series at each crossing whose job is to be sharply nonlinear: nearly non-conducting below a threshold voltage so it blocks the sneak paths, then snapping to a conducting state above threshold so it passes the cell you are addressing.

Undoped vanadium sesquioxide is metallic at room temperature; its intrinsic Mott metal-insulator transition is cryogenic, near 150 to 170 kelvin. The room-temperature insulating state a selector needs exists only because the chromium doping opens a paramagnetic insulating phase, so the dopant is load-bearing rather than cosmetic. The obvious guess is still that the switch is that electronic transition, but the authors argue otherwise, and the distinction matters. They report that their data fit a thermal-runaway picture better: the one-time forming step builds a conductive filament through the film, titanium-doped through the interfacial layer in the crystalline case and chromium-doped through the whole five nanometres in the amorphous case, and within that filament self-heating drives an electrothermal negative-differential-resistance switch. The switch is volatile, relaxing back to the blocking state when the drive is removed, which is what a selector wants. Tracing the behaviour to that specific interfacial layer and to titanium migrating in from the electrode is the paper's real contribution: it turns a switch that survives to five nanometres into a process knob, rather than the fragile collective transition that thinning usually destroys.¹

One property of this class of device matters more than any other for what follows. The switch is designed to be nonlinear and thresholded. It is not a wire and not a linear resistor; below threshold it is meant to pass almost nothing, and its usefulness is precisely that it distorts the current-voltage relationship. Hold that thought.

Where a skeptic should push

The most load-bearing weakness is that the paper's best property leans on its least controllable feature. A forming step is required in both crystalline and amorphous films, and the improved leakage and abruptness are attributed to an interfacial amorphous layer plus titanium interdiffusion. Forming is a one-time electrical conditioning that is notoriously variable from device to device, and pinning a key parameter to a self-organised interface reaction is a uniformity and yield risk, not a reassurance. Demonstrated here is the behaviour of a small number of devices, electrically and under the microscope. Not demonstrated is what a large array of them does: cycle-to-cycle variability, endurance, and above all the spread of forming voltages and thresholds across thousands of sites. The authors are candid that the origin of the improved behaviour is not fully settled, and the decisive interfacial imaging comes from thicker companion films of ten and one hundred and twenty nanometres rather than direct microscopy of the five-nanometre devices that were electrically tested, so the mechanism is a plausible inference stitched to the electrical curves rather than a closed case. Two process facts compound the worry: the crystalline films were grown near six hundred degrees, which is incompatible with building the selector on top of finished readout electronics, and the interfacial layer and titanium diffusion the paper leans on are by-products of that hot process, not a knob the fab sets.

One fact I could not pin down is load-bearing, so I flag it: the paper never states the measurement temperature. The mechanism it does settle, and it settles it against the easy assumption: the authors explicitly favour electrothermal thermal-runaway in the formed filament over the electronic Mott transition, and they report no cooling, so room-temperature operation is a strong inference. That resolves half the array question and sharpens the other half. A warm, usable selector is the charitable case for any device meant to sit near tissue, but an electrothermal switch buys its abruptness with self-heating, and self-heating makes thermal crosstalk between neighbours the scaling limit in a dense array, exactly where you would want the density. And because the switch lives in a formed filament, its uniformity is a filament-formation problem, which is about as variable as device physics gets.

Why the recording array pays for a transistor

A high-density microelectrode array has the same wiring problem this selector is built for: how to reach far more electrodes than you have amplifier channels or wires. There are three answers in practice, not two. The amplifier-per-pixel design puts an amplifier under every electrode. The dominant high-density design is switch-matrix multiplexed: a transistor or two at each site handles routing while amplifiers and analogue-to-digital converters are shared and time-multiplexed across many electrodes, which is how the large CMOS arrays put tens of thousands of electrodes on a chip and read a selected subset at once. The third answer is the passive crossbar: row and column lines, a device per crossing, no transistor, cheap and dense but limited by sneak paths, and rare for high-fidelity recording precisely because of them. A good two-terminal selector is what would rescue that third option, and on paper a five-nanometre, low-leakage selector reads as a blueprint for a transistor-free electrode-addressing fabric. That is the tempting reading, and it is wrong for the part of the array that matters most.

The load-bearing reason needs stating carefully, because the intuitive version is wrong and a good reviewer will catch it. The intuitive version says a spike of tens to low hundreds of microvolts never reaches this selector's one to one and a half volt threshold, so the device stays off and blocks the signal outright. That is not what happens. A recording amplifier presents a very high input impedance, so the open-circuit voltage across a non-conducting selector still appears at its input, and you could even hold the device just below threshold to give it a small-signal operating point. The spike is not annihilated, and I withdraw the stronger claim that it is. The real disqualification is worse and, unlike the blocking story, it cannot be biased away. In its off state the selector is a gigaohm-class series impedance, and a gigaohm resistor carries Johnson noise of order four microvolts per root hertz, which across a ten-kilohertz spike band is roughly four hundred microvolts, several times a spike and far above the few-microvolt floor a good front end reaches. Bias it up toward threshold to lower that impedance and you enter the negative-differential-resistance regime, where random-telegraph noise, excess low-frequency noise, and metastability are at their worst, so the noise gets worse, not better. On top of that the device is deliberately nonlinear, so it intermodulates the very waveform you are trying to preserve. Any one of these is disqualifying for a linear microvolt path; together they make a two-terminal threshold selector the wrong primitive to sit between an electrode and its amplifier. The corrected argument is not that the threshold blocks the spike, but that the device's impedance, noise, and nonlinearity destroy it.

The paper's own findings sharpen a second objection that applies even where a switch would be acceptable. Forming is an electrical stress you would have to apply to a device sitting in series with a delicate electrode-electrolyte interface, and the per-site variability of forming voltage and threshold becomes per-channel offset variability across thousands of sites, a calibration burden the active-pixel design simply does not incur. And the mechanism the microscopy uncovered, an interfacial layer and titanium diffusing from the electrode, is a reminder that any such device at the tissue boundary introduces materials with their own corrosion, drift, and biocompatibility questions in saline at body temperature, and an electrothermal switch additionally dumps heat at the site where the tissue least tolerates it.

Where the device genuinely could map is the other direction of the array. Selection is a real need on the stimulation side, where you route volt-scale pulses to chosen electrodes, and in actively-interrogated impedance modes, where you drive a known current and can live with, even exploit, a threshold. There a compact, low-leakage, thin selector could enable denser transistor-free stimulus or impedance routing, provided forming and uniformity are solved. So the honest implication is a split verdict: the opportunity is on the stimulate-and-interrogate fabric, and the threat is the temptation to read selector scaling as a road to denser recording arrays. It is not, because the recording bottleneck was never wire count. State-of-the-art arrays already field tens of thousands of electrodes; what binds them is per-channel input-referred noise against the electrode impedance, the number of channels you can digitise at once through shared readout, and the power and heat you can spend in tissue. A selector attacks wire count, the one axis that was not limiting, and it helps none of the three that are.

The bottom line

Established, at the level of a few devices: chromium-doped vanadium-oxide selector behaviour survives and improves at a five-nanometre thickness, with the improvement traced to a specific interfacial mechanism, at the cost of a required forming step and a dependence on an uncontrolled interface layer. Settled: the switch is electrothermal, which favours room-temperature use, though the paper never states the measurement temperature. Not shown: array-scale uniformity, endurance, and the spread of filament formation across many devices. My extrapolation to microelectrode arrays lands on a hard boundary: for microvolt recording a threshold selector in the sense path is contraindicated not because it blocks the spike but because its gigaohm off-state impedance, its added noise, and its nonlinearity would swamp it, and its plausible home is instead stimulation and impedance routing. What would confirm or break the array case is array-level forming-yield and uniformity data, an explicit operating temperature, and, for anyone still tempted by the recording idea, a measurement that the device adds negligible noise and distortion to a microvolt signal, which its off-state impedance and nonlinearity make unlikely.