Reading a single skyrmion electrically: a picoamp tunnelling measurement, from idea to verdict.

In a normal magnet, neighbouring spins point the same way and electrons tunnel through two clean spin channels. Inside a skyrmion the spins twist, the channels mix, and the local density of electronic states changes. A scanning-tunnelling microscope can see that change in the differential conductance, dI/dV, at a single atom.

What NCMR means. NCMR stands for non-collinear magnetoresistance. Spins are collinear when they all lie along one line, pointing parallel or antiparallel, as in an ordinary ferromagnet. They are non-collinear when neighbouring spins sit at an angle to each other, as they do all the way round a skyrmion. NCMR is the change in tunnelling resistance caused by that angle between neighbouring spins.

• Stack 22k (about one skyrmion per dot) and 22d (four to five) were grown on 4-inch wafers; magnetic force microscopy confirmed skyrmions in the film
• Barrier candidates were screened by resistance: Al₂O₃ chosen, Ta₂O₅ rejected as leaky
• First batch: about 300 dots of 250 to 400 nm with a 5 nm Al₂O₃ barrier

The MTJ: compare two magnets

A magnetic tunnel junction sandwiches a thin insulator between two ferromagnets. One, the reference layer, is pinned; the other, the free layer, can be flipped. Electrons tunnel more easily when the two magnetisations are parallel than when they are antiparallel, so the resistance reads the bit. This tunnelling magnetoresistance is the read mechanism of today’s MRAM and hard-disk heads, and it needs the second, pinned magnet as a reference.

The NCMR device: sense the twist inside one magnet

Non-collinear magnetoresistance (NCMR) needs only one magnetic layer. Where neighbouring spins are parallel, electrons see two clean spin channels. Where they twist, as they do through a skyrmion, spin-orbit coupling mixes the channels and the local density of electronic states changes. The tunnelling conductance, dI/dV, changes with it, even through a non-magnetic top electrode. The bit is not the direction of a layer but whether a skyrmion is there.

In this project, each nanodot was one bit: a dot holding a skyrmion is a 1, an empty dot a 0. The measurement asked whether dI/dV could tell the two apart electrically.

The bits were written with the field’s hysteresis. To write a 1, the field came down to the measurement point from above, from positive saturation; to write a 0, it came up from below, from negative field. At the same field the dot could then hold either state, so a 1 and a 0 could be compared without the field itself changing the reading.

What the devices are for

The long-term goal is non-volatile data storage and logic. Skyrmions are tens of nanometres across, topologically stable, and can be moved by small currents, which makes them candidates for racetrack memory: bits stored as a train of skyrmions along a magnetic track, shifted past a fixed read head. They are also studied for logic and neuromorphic computing.

Any such memory needs a simple electrical read-out of a single skyrmion. NCMR could provide it with a simpler stack than an MTJ, because the read head needs no pinned reference magnet. That is what this project set out to test: whether the effect seen at a single atom under a microscope survives in a wired device averaged over a whole dot.

I adapted the circuit microscopists use for tunnelling spectroscopy. A small AC excitation at 77.77 Hz rides on a DC bias; a bias adder combines the two and applies them to the device. The current through the device goes to a low-noise current preamplifier and a lock-in amplifier, which picks out the part at 77.77 Hz. That ratio of AC current to AC voltage is dI/dV at the chosen bias. A separate voltmeter reads the DC bias actually reaching the device.

Everything around the device mattered as much as the circuit:

• a shielded NCMR cell with LEMO connectors, an EMI shield and a 44-pin chip-carrier stage, built in two versions for out-of-plane and in-plane fields
• a holder that fitted the cell into a 0–1 T electromagnet
• wire bonding set up in the lab, with a tested bonding recipe
• a grounding scheme that removed ground loops between the source, voltmeter and lock-in
• a LabVIEW program that ran bias and field sweeps and logged every run, through 29 versions

I also ran a small designed experiment on the lock-in itself, comparing sensitivity, reserve, time constant, filter slope and the line filters, and kept the combination that gave the most stable reading.

A 100 kΩ resistor gave a flat dI/dV that did not change with DC bias, exactly as a linear part should, with noise around 1 nV and the phase near zero.

A diode was the real test, because its resistance falls by orders of magnitude as the bias rises. I measured its I-V curve with a separate DC setup, fitted it, and calculated what dI/dV should be. The lock-in measurement matched the model at currents below 100 nA and across device resistances from under 1 kΩ to 8,000 kΩ.

Above 0.38 V the two disagreed. I traced it to two causes in the circuit, not the method: a voltage-divider effect once the diode's resistance fell toward the 10 Ω sense resistor, and the source hitting its 200 mA current limit. The fix was to measure the circuit's transfer function for both AC and DC, so the bias at the device is known rather than assumed.

A tunnel-junction test device then gave 2.37 µS from its I-V curve and from dI/dV, the same answer two independent ways.

Along the way I checked that the voltmeter would not load a high-resistance device (it needs well over 100 MΩ input resistance) and found why several test junctions had died: probing them with a handheld multimeter pushed too much current through the barrier. From then on, resistance checks used a source-meter at 0.05 V.

The answer was not the one everyone hoped for, and that is still a result. Because the instrument had been proven step by step, the switching could be pinned on the device, not the tester, and the next batch could be planned on evidence.

The first question was whether the signal came from tunnelling at all, or from leakage in the wiring or across the dot. With the device and chip carrier removed, the background was about 1 nS and flat with bias. With a device in place the signal ranged from about 1 to 10 nS and rose non-linearly with bias, which only tunnelling explains.

Then I mapped how far the devices could be pushed. Below 1 V, dI/dV was reproducible. Above 2 V the barrier broke down, temporarily at first and then permanently. Above 0.5 V the signal drifted by about 1 nS, ten times the measurement's noise floor of under 0.1 nS, which pointed to the barrier starting to break down locally.

To reach higher bias without damage, I added a pulsed bias mode, applying the bias for 2 to 20 seconds at a time, which reached 2 V on the device. I also upgraded the DC readout with a current preamplifier until it resolved changes under 0.5 pA, and set a target of 0.01 nS for the next stage.

Bias sweeps at fields from +0.3 to −0.3 T, and field sweeps between ±0.5 T, showed no change that could be pinned on a skyrmion. Within this window, any skyrmion signal was smaller than the 1 nS drift.

Under steady bias, the 400 nm dots switched on their own between two current levels, about 1 pA apart, even at zero field. A skyrmion hopping in and out of the dot would look exactly like this, so it was tempting to claim it. Instead I tested every other explanation first:

• The setup? No switching with no device in place. Ruled out.
• One bad dot? It appeared in several 400 nm dots. Ruled out.
• Any tunnel junction? It did not appear in plain magnetic tunnel junctions. Ruled out.
• Size? 300 nm dots needed higher bias and switched less often.
• Magnetism? A 0.5 T field, strong enough to wipe out skyrmions, did not stop it, and it did not repeat reliably in the same dot.

I then compared the step size with the literature on oxide breakdown. About 1 pA through a 400 nm dot is roughly 1,000 µA/cm², the same order as random telegraph noise from charge traps in thin silicon oxide after soft breakdown. Together, the evidence pointed to traps in the Al₂O₃ barrier, not skyrmions.

The first batch could not show a skyrmion signal, and I said so plainly in the weekly reports instead of over-reading the switching. What the work did deliver was a validated instrument, a map of what the devices could survive, and specific changes for the next batch:

• a barrier change: a thicker Al₂O₃ or MgO once the deposition tool was back, to widen the safe bias window
• a measurement plan ranked by what would work: DC conductance first, then pulsed-bias dI/dV, with resolution pushed from 0.1 nS to 0.01 nS
• a direct test of the oxide explanation: switching recorded at 0, 0.1, 0.5 and 1 T for an hour each, with step size analysed against bias and current
• device handling rules that stopped junctions being destroyed during checks

Behind it sat a MATLAB pipeline I wrote to import thousands of sweeps, filter them by bias or field, bin and average them, and plot each run the same way, so every report compared like with like.