A magnetic tunnel junction that is a sensor, a memory, and a switch

What the work claims

Magnetic tunnel junctions (MTJs) and memristors are usually competing technologies: the first, an insulating MgO barrier between magnetic layers, is prized for high-yield sensors and magnetic memory; the second, a soft dielectric breakdown that leaves a reversible conductive filament, is prized for dense analog storage and neuromorphic crossbars. Prior attempts to combine them typically lost either the magnetoresistance or most of the resistive switching ratio. This work demonstrates bipolar resistive switching coexisting with an intact tunneling magnetoresistance in the same MgO barrier, without degrading either property: the devices show a TMR of about 45%, equivalent to the fabricated magnetic sensors without memristive functionality, while also switching non-volatilely and quasi-analogically under electrical pulses down to nanosecond width.¹

The boldest sub-claim is reconfigurability: switching the junction to its low-resistance state completely suppresses its TMR, because the filament shunts the tunnel barrier, and switching back restores the original field response with no apparent degradation, repeatedly. A magnetic sensor that can be electrically taken out of the circuit and put back in is a new circuit primitive, and the authors name the application themselves: reprogrammable multifunctional field sensor arrays and reconfigurable neuromorphic circuits whose topology can be modified after fabrication.

How it works

The devices are cylindrical MgO MTJs with a soft pinned sensing layer, fabricated at diameters from 5 to 100 um, the geometry used for ultrasensitive field sensors. A three-step annealing protocol linearizes the magnetic transfer curve: the presented device shows a linear response to applied field with no hysteresis over a range of about 60 Oe, the transfer characteristics wanted for sensing, with TMR of about 45% at 0.1 mA bias current in a 55 um junction. The same stack then undergoes electroforming: a first, higher positive bias creates a soft breakdown, after which the barrier switches reversibly between a high-resistance state (HRS) and a low-resistance state (LRS), with set onset near 0.7 V and reset onset near 0.85 V. Reads are non-invasive 50 mV pulses taken 10 ms after each write; resistance-versus-write-voltage loops cycled 200 times show the full hysteresis, and retention holds both states for about 1e4 s under a constant 100 mV read, with the HRS current-voltage curve fitting the Simmons tunneling model.¹

Two results carry the physics. First, applying a saturating magnetic field has no observable effect on the memristive switching, and cycling the resistance states does not degrade the magnetic response: the two effects are genuinely independent. Second, the TMR does not just drop in the LRS, it vanishes entirely, and it returns after reset to its original value. A two-channel parallel-conduction model explains both facts: spin-polarized electrons tunnel through the barrier in one channel, giving the field-dependent response, while the filament is a spin-independent shunt resistance in parallel. The measured TMR tracks the resistance-area product linearly across quasi-analog intermediate states, exactly the signature of that model. Those intermediate states are accessible by partial reset, sweeping the reverse stop voltage from -1 to -1.4 V in 0.01 V steps across 40 curves, and each stable intermediate resistance, with its correspondingly scaled TMR, is non-volatile.¹

The energy figures make the device practical: with a Ta-doped MgO barrier, electroforming consumes about 20% less energy at the cost of TMR falling to 25%, and in the Ta-interlayer stack written with 10 ns pulses, forming takes about (350 +/- 50) pJ while set and reset cost about 5 pJ and 100 pJ per operation. Smaller junctions need less forming energy because the smaller, more resistive device limits total heating.¹

Where a skeptic should push

The load-bearing assumption is that filamentary switching can be made repeatable enough to carry an analog value, and the present data says only that it has survived 200 cycles on studied devices. That is a feasibility demonstration, not a reliability qualification: filamentary memristors are notorious for cycle-to-cycle and device-to-device variation, retention of intermediate states is shown only to about 1e4 s, and nothing approaching an endurance spec for quasi-analog weights is offered. The honest reading is that the coexistence is proven and the engineering maturity is early.

Second, everything here is measured on 5 to 100 um research-grade pillars. Industrial MTJs are sub-micron and MRAM-qualified; these are four to five orders of magnitude away from that density, and the memristive behaviour at scaled dimensions is extrapolation, though the area-dependence data (smaller devices form more efficiently) points the right way. Third, the doping trade is blunt: the 20% forming-energy saving costs TMR of 45% down to 25%, so the power-optimised stack is a worse sensor, and the paper does not resolve that coupling. None of this overturns the result; all of it bounds how far the numbers travel from the wafer they were measured on.

When one junction is sensor, memory, and switch

For microelectrode array instrumentation, the interesting question is not whether this specific stack belongs in an array. It is that three functions the acquisition chain currently buys in separate components can, in principle, share one CMOS-compatible element. Consider what a high-density MEA system pays for today. It pays for magnetic field cleanliness twice over: stimulation and magnetic actuation generate field transients that couple into high-impedance neural amplifiers, so rigs buy shielding and distance, and rarely measure the offending field at all. It pays for non-volatile calibration storage per channel or per column, streaming coefficients off-chip through the same ADC budget that the biology competes for. And it pays, increasingly, for analog in-memory compute at the back end to beat the egress wall. This paper shows a junction that does all three: a linear, hysteresis-free field sensor over about 60 Oe with 45% TMR; a non-volatile quasi-analog weight stable at tens of intermediate states; and a switch that can electrically sever itself from the signal path, or rejoin it, on demand.

The non-obvious opportunity is provenance. The severable-sensor trick the authors demonstrate, shorting one leg of a Wheatstone bridge to turn it into a gradiometer, transposes directly to instrumentation: an array could carry its own field metrology, logging the ambient field in which each recording block was taken, with the log stored in the same junctions that sensed it. Calibration state and the conditions under which that calibration was valid would live in one physical object, at picojoule write cost. For a field where arguments about stimulation artifacts, grounding topology and 50 or 60 Hz pickup are settled by anecdote, self-reporting instruments are a genuine capability shift.

The threats are equally concrete. A filament-forming element sitting near a neural front end is a soft-breakdown device by design, and its cycle-to-cycle variation, 1e4 s retention, and TMR-versus-power trade are precisely the failure modes an instrumentation buyer cannot accept at spec today. More subtly, a magnetic sensor integrated into the readout die will pick up the fields of its own stimulation currents, so the artifact monitor and the artifact source share a board; used naively, the sensor manufactures the interference it reports. And the reconfigurability that makes the device attractive, post-fabrication topology changes by write pulses, is also a reliability and security surface: an instrument whose connectivity can be rewritten by its control bus must defend that bus as part of its measurement validity. The paper supplies the device physics; the acquisition chain would supply the discipline.

The bottom line

Established, on measured devices: MgO MTJs can host bipolar, non-volatile, quasi-analog resistive switching with set and reset at roughly 0.7 and 0.85 V while retaining a 45% linear, hysteresis-free magnetoresistance; the spintronic response can be fully suppressed in the low-resistance state and recovered on reset, with the two-channel parallel-conduction model fitting the data across intermediate states; and 10 ns writes at about 5 pJ per set are demonstrated in a Ta-doped stack at 20% lower forming energy with TMR reduced to 25%. Not established: endurance beyond a few hundred cycles, retention beyond about 1e4 s, device-to-device uniformity, or behaviour at production-scale dimensions. What would confirm the platform is cycle-count and array-level uniformity data on scaled junctions; what would break it is evidence that filament statistics at sub-micron dimensions destroy either the analog state stability or the recovered TMR. For array hardware, the lasting contribution is conceptual: it collapses three back-end line items, field sensing, non-volatile analog storage, and reconfigurable connectivity, into one junction, and hands instrument designers a new question worth asking of every new device: which of my separate components could this replace, and what does the merge cost in reliability?