Spin-Transfer Torque Magnetic Tunnel Junctions

• Spin-transfer torque magnetic tunnel junctions are devices that use spin-polarized currents to switch magnetic layers and read out states via tunnel magnetoresistance.
• They encompass a range of architectures—from MgO-based to fully insulating and ferrimagnetic systems—where engineering interfaces and layers tune torque components and bias-dependent behavior.
• STT-MTJs enable applications in memory, computing, and microwave oscillators, with performance driven by material selection, interfacial resonances, and thermal as well as spin-diffusion effects.

Spin-transfer torque magnetic tunnel junctions (STT-MTJs) are tunnel devices in which a spin-polarized current crossing an ultrathin barrier exerts torque on a magnetically soft or otherwise switchable layer, while the same junction provides electrical readout through tunnel magnetoresistance (TMR). In the canonical implementation, a free magnetic layer and a reference layer are separated by MgO or a related insulator, but the research literature now spans perpendicular and in-plane ferromagnetic MTJs, resonant double-barrier structures with embedded nanoparticles, ferrimagnetic and antiferromagnetic tunnel junctions, fully insulating junctions, continuous-free-layer logic elements, and oscillator-oriented nanocontacts. Across these variants, performance is set by the coupled behavior of TMR, damping-like and field-like torque, interfacial symmetry filtering, resonant transmission, spin diffusion, thermal activation, and bias-driven nonlinearity (Useinov et al., 2016, Leutenantsmeyer et al., 2013, Kang et al., 3 Sep 2025).

1. Core physical framework

STT-MTJs combine two reciprocal effects. First, the junction resistance depends on the relative alignment of the magnetic electrodes, giving the standard TMR measures

$$\mathrm{TMR}(V)=\frac{R_{\mathrm{AP}}(V)-R_{\mathrm{P}}(V)}{R_{\mathrm{P}}(V)}$$

or, equivalently in conductance form,

$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$

Second, the spin current crossing the barrier transfers angular momentum to the free layer. In the usual Landau–Lifshitz–Gilbert–Slonczewski description,

$$\frac{d\mathbf{m}}{dt}=-\gamma\,\mathbf{m}\times\mathbf{H}_{\mathrm{eff}}+\alpha\,\mathbf{m}\times\frac{d\mathbf{m}}{dt}+\tau_{\mathrm{STT}},$$

with damping-like and field-like components

$$\tau_{\parallel}=\frac{\hbar}{2e}\frac{I}{M_s t}g(\theta)\,\mathbf{m}\times(\mathbf{m}\times\mathbf{p}), \qquad \tau_{\perp}=\frac{\hbar}{2e}\frac{I}{M_s t}h(\theta)\,\mathbf{m}\times\mathbf{p}.$$

In conventional symmetric metallic MTJs, the angular dependence is often approximately proportional to $\sin\theta$, and the standard low-bias expansion is

$$T_{\parallel}(V)=a_1V+a_2V^2,\qquad T_{\perp}(V)=b_0+b_2V^2,$$

but several later results show that this is only a limiting case (Useinov et al., 2016, Manchon et al., 2012, Wang et al., 2010).

A recurrent misconception is that the in-plane torque is the only practically relevant component. That approximation is often computationally convenient, yet direct time-domain ST-FMR measurements showed that the field-like component can become substantial at application-relevant bias, with a maximal $\tau_\perp$ corresponding to an effective field of about $30$ Oe, and that the torque vector departs markedly from simple lowest-order Taylor expansions at high $|V|$ (Wang et al., 2010). A second misconception is that the low-bias symmetry pattern is universal: in fact, diffusion, resonant transport, or fully insulating electrodes can qualitatively alter both the magnitude and the parity of the torque terms (Manchon et al., 2012).

2. Materials systems and stack engineering

The mainstream STT-MTJ platform remains the MgO-based ferromagnetic junction, implemented either with in-plane anisotropy or with perpendicular magnetic anisotropy (PMA) from CoFeB/MgO interfaces. Perpendicular Co-Fe-B/MgO junctions with Ta(15)/Co-Fe-B(1.0)/MgO(0.84)/Co-Fe-B(1.2)/Ta(5.0)/Ru(3.0) showed TMR ratios up to $64\%$ at a $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$0 monolayer tunnel barrier thickness, while patterned circular junctions in the nominal $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$1–$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$2 nm range were used for low-current DC-STT studies (Leutenantsmeyer et al., 2013). In a different PMA setting, circular CoFeB free layers of diameter $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$3 nm and $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$4 nm with $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$5 nm, $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$6 A/m, $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$7 J/m, and $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$8 J/m$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$9 provided the micromagnetic basis for write-error-rate studies in nanoscale p-MTJs (Das et al., 13 Nov 2025).

Beyond CoFeB, Heusler and ferrimagnetic electrodes have been investigated to raise spin polarization or modify damping. Polycrystalline B2-type Co$$\frac{d\mathbf{m}}{dt}=-\gamma\,\mathbf{m}\times\mathbf{H}_{\mathrm{eff}}+\alpha\,\mathbf{m}\times\frac{d\mathbf{m}}{dt}+\tau_{\mathrm{STT}},$$0FeAl MTJs on amorphous Si/SiO$$\frac{d\mathbf{m}}{dt}=-\gamma\,\mathbf{m}\times\mathbf{H}_{\mathrm{eff}}+\alpha\,\mathbf{m}\times\frac{d\mathbf{m}}{dt}+\tau_{\mathrm{STT}},$$1, enabled by an MgO buffer, reached TMR up to $$\frac{d\mathbf{m}}{dt}=-\gamma\,\mathbf{m}\times\mathbf{H}_{\mathrm{eff}}+\alpha\,\mathbf{m}\times\frac{d\mathbf{m}}{dt}+\tau_{\mathrm{STT}},$$2 for a CFA/MgO/CoFe structure on a $$\frac{d\mathbf{m}}{dt}=-\gamma\,\mathbf{m}\times\mathbf{H}_{\mathrm{eff}}+\alpha\,\mathbf{m}\times\frac{d\mathbf{m}}{dt}+\tau_{\mathrm{STT}},$$3 nm MgO buffer. In the corresponding STT nanopillars with a $$\frac{d\mathbf{m}}{dt}=-\gamma\,\mathbf{m}\times\mathbf{H}_{\mathrm{eff}}+\alpha\,\mathbf{m}\times\frac{d\mathbf{m}}{dt}+\tau_{\mathrm{STT}},$$4 nm CFA switching layer, the intrinsic critical current density was $$\frac{d\mathbf{m}}{dt}=-\gamma\,\mathbf{m}\times\mathbf{H}_{\mathrm{eff}}+\alpha\,\mathbf{m}\times\frac{d\mathbf{m}}{dt}+\tau_{\mathrm{STT}},$$5 A/cm$$\frac{d\mathbf{m}}{dt}=-\gamma\,\mathbf{m}\times\mathbf{H}_{\mathrm{eff}}+\alpha\,\mathbf{m}\times\frac{d\mathbf{m}}{dt}+\tau_{\mathrm{STT}},$$6, and ferromagnetic resonance gave a Gilbert damping constant of about $$\frac{d\mathbf{m}}{dt}=-\gamma\,\mathbf{m}\times\mathbf{H}_{\mathrm{eff}}+\alpha\,\mathbf{m}\times\frac{d\mathbf{m}}{dt}+\tau_{\mathrm{STT}},$$7, nearly independent of CFA thickness from $$\frac{d\mathbf{m}}{dt}=-\gamma\,\mathbf{m}\times\mathbf{H}_{\mathrm{eff}}+\alpha\,\mathbf{m}\times\frac{d\mathbf{m}}{dt}+\tau_{\mathrm{STT}},$$8 to $$\frac{d\mathbf{m}}{dt}=-\gamma\,\mathbf{m}\times\mathbf{H}_{\mathrm{eff}}+\alpha\,\mathbf{m}\times\frac{d\mathbf{m}}{dt}+\tau_{\mathrm{STT}},$$9 nm (1408.0341). This is notable because the same study attributed the reduced $$\tau_{\parallel}=\frac{\hbar}{2e}\frac{I}{M_s t}g(\theta)\,\mathbf{m}\times(\mathbf{m}\times\mathbf{p}), \qquad \tau_{\perp}=\frac{\hbar}{2e}\frac{I}{M_s t}h(\theta)\,\mathbf{m}\times\mathbf{p}.$$0 primarily to the lower damping of the polycrystalline CFA free layer relative to the epitaxial case.

Ferrimagnetic Mn$$\tau_{\parallel}=\frac{\hbar}{2e}\frac{I}{M_s t}g(\theta)\,\mathbf{m}\times(\mathbf{m}\times\mathbf{p}), \qquad \tau_{\perp}=\frac{\hbar}{2e}\frac{I}{M_s t}h(\theta)\,\mathbf{m}\times\mathbf{p}.$$1Ga introduces a different regime. First-principles NEGF+DFT calculations for Fe/MgO/Mn$$\tau_{\parallel}=\frac{\hbar}{2e}\frac{I}{M_s t}g(\theta)\,\mathbf{m}\times(\mathbf{m}\times\mathbf{p}), \qquad \tau_{\perp}=\frac{\hbar}{2e}\frac{I}{M_s t}h(\theta)\,\mathbf{m}\times\mathbf{p}.$$2Ga and Mn$$\tau_{\parallel}=\frac{\hbar}{2e}\frac{I}{M_s t}g(\theta)\,\mathbf{m}\times(\mathbf{m}\times\mathbf{p}), \qquad \tau_{\perp}=\frac{\hbar}{2e}\frac{I}{M_s t}h(\theta)\,\mathbf{m}\times\mathbf{p}.$$3Ga/MgO/Mn$$\tau_{\parallel}=\frac{\hbar}{2e}\frac{I}{M_s t}g(\theta)\,\mathbf{m}\times(\mathbf{m}\times\mathbf{p}), \qquad \tau_{\perp}=\frac{\hbar}{2e}\frac{I}{M_s t}h(\theta)\,\mathbf{m}\times\mathbf{p}.$$4Ga found long-range spatial oscillations of the STT extending tens of angstroms into Mn$$\tau_{\parallel}=\frac{\hbar}{2e}\frac{I}{M_s t}g(\theta)\,\mathbf{m}\times(\mathbf{m}\times\mathbf{p}), \qquad \tau_{\perp}=\frac{\hbar}{2e}\frac{I}{M_s t}h(\theta)\,\mathbf{m}\times\mathbf{p}.$$5Ga, with the oscillation wave number governed mainly by the longitudinal lattice constant $$\tau_{\parallel}=\frac{\hbar}{2e}\frac{I}{M_s t}g(\theta)\,\mathbf{m}\times(\mathbf{m}\times\mathbf{p}), \qquad \tau_{\perp}=\frac{\hbar}{2e}\frac{I}{M_s t}h(\theta)\,\mathbf{m}\times\mathbf{p}.$$6 rather than by barrier thickness or interface spacing (Stamenova et al., 2020). This differs sharply from the usual short-range decay picture in conventional ferromagnets. A plausible implication is that ferrimagnetic free layers can no longer be treated as purely interfacial torque absorbers.

The range of electrode types extends even further. Fully insulating MTJs, in which both magnetic electrodes are insulators, were predicted to exhibit transport dominated by evanescent states and spin-dependent Fowler–Nordheim tunneling, with an out-of-plane torque that generally dominates the in-plane torque and is symmetric at low bias (Manchon, 2012). At the opposite end of magnetic order, all-antiferromagnetic PtMn$$\tau_{\parallel}=\frac{\hbar}{2e}\frac{I}{M_s t}g(\theta)\,\mathbf{m}\times(\mathbf{m}\times\mathbf{p}), \qquad \tau_{\perp}=\frac{\hbar}{2e}\frac{I}{M_s t}h(\theta)\,\mathbf{m}\times\mathbf{p}.$$7MgO$$\tau_{\parallel}=\frac{\hbar}{2e}\frac{I}{M_s t}g(\theta)\,\mathbf{m}\times(\mathbf{m}\times\mathbf{p}), \qquad \tau_{\perp}=\frac{\hbar}{2e}\frac{I}{M_s t}h(\theta)\,\mathbf{m}\times\mathbf{p}.$$8PtMn$$\tau_{\parallel}=\frac{\hbar}{2e}\frac{I}{M_s t}g(\theta)\,\mathbf{m}\times(\mathbf{m}\times\mathbf{p}), \qquad \tau_{\perp}=\frac{\hbar}{2e}\frac{I}{M_s t}h(\theta)\,\mathbf{m}\times\mathbf{p}.$$9 tunnel junctions were shown to support room-temperature TMR of $\sin\theta$0 and current-induced switching at current densities of the order of $\sin\theta$1 MA/cm$\sin\theta$2, establishing that the STT-MTJ concept is not confined to ferromagnets (Kang et al., 3 Sep 2025).

3. Angular dependence, bias dependence, and microscopic torque generation

In many descriptions, the in-plane torque is parameterized by Slonczewski’s form

$\sin\theta$3

where $\sin\theta$4 measures angular asymmetry. Ordered CoFe/Mg-B-O/CoFe(001) junctions provide a clear example of how far real devices can depart from the symmetric $\sin\theta$5 limit. For CoFe/Mg$\sin\theta$6BO$\sin\theta$7(3L)/MgO(3L)/Mg$\sin\theta$8BO$\sin\theta$9(3L)/CoFe, $$T_{\parallel}(V)=a_1V+a_2V^2,\qquad T_{\perp}(V)=b_0+b_2V^2,$$0; for CoFe/Mg$$T_{\parallel}(V)=a_1V+a_2V^2,\qquad T_{\perp}(V)=b_0+b_2V^2,$$1BO$$T_{\parallel}(V)=a_1V+a_2V^2,\qquad T_{\perp}(V)=b_0+b_2V^2,$$2(5L)/MgO(3L)/Mg$$T_{\parallel}(V)=a_1V+a_2V^2,\qquad T_{\perp}(V)=b_0+b_2V^2,$$3BO$$T_{\parallel}(V)=a_1V+a_2V^2,\qquad T_{\perp}(V)=b_0+b_2V^2,$$4(5L)/CoFe, $$T_{\parallel}(V)=a_1V+a_2V^2,\qquad T_{\perp}(V)=b_0+b_2V^2,$$5; whereas CoFe/Mg$$T_{\parallel}(V)=a_1V+a_2V^2,\qquad T_{\perp}(V)=b_0+b_2V^2,$$6BO$$T_{\parallel}(V)=a_1V+a_2V^2,\qquad T_{\perp}(V)=b_0+b_2V^2,$$7(3L)/MgO(3L)/Mg$$T_{\parallel}(V)=a_1V+a_2V^2,\qquad T_{\perp}(V)=b_0+b_2V^2,$$8BO$$T_{\parallel}(V)=a_1V+a_2V^2,\qquad T_{\perp}(V)=b_0+b_2V^2,$$9(3L)/CoFe gives an almost symmetric curve with $\tau_\perp$0. Disorder in the B distribution suppresses the skewness, driving $\tau_\perp$1 in the disordered Mg$\tau_\perp$2BO$\tau_\perp$3 case and to about $\tau_\perp$4 in the disordered Mg$\tau_\perp$5BO$\tau_\perp$6 case (Tang et al., 2016). The underlying mechanism is interfacial resonance: B diffusion turns interfacial MgO layers into conductive Mg-B-O, and when the remaining undoped MgO is thin enough, resonant $\tau_\perp$7 hot spots dominate both conductance and torque.

A persistent misconception is that torque asymmetry is purely a materials-polarization effect. The B-doped MgO results show instead that ordering, hybridization, and barrier topology can be decisive. B only in the middle of the barrier yields $\tau_\perp$8, whereas B at the interfaces can raise $\tau_\perp$9 to the $30$0–$30$1 range (Tang et al., 2016). This suggests that angular skewness is better viewed as a transport-geometry property than as a fixed electrode constant.

Bias dependence is equally non-universal. A ballistic-plus-diffusive theory of metallic MTJs showed that spin diffusion in the electrodes mixes the transverse spin-current components, so that even if the injected interfacial torque obeys the conventional pattern $30$2 and $30$3, the effective field-like torque acquires a linear term,

$30$4

with

$30$5

where $30$6 and $30$7 (Manchon et al., 2012). The same framework predicts non-conventional thickness dependence when the free-layer thickness becomes comparable to $30$8 or $30$9.

Direct time-domain ST-FMR at high bias provided the experimental counterpart. In MgO MTJs with RA about $|V|$0 $|V|$1m$|V|$2 and TMR $|V|$3–$|V|$4, $|V|$5 became strongly asymmetric, being $|V|$6–$|V|$7 larger at high negative bias than at high positive bias, while $|V|$8 saturated so that $|V|$9 crossed over from quadratic to approximately linear in $64\%$0 (Wang et al., 2010). This directly contradicts the widespread practice of extrapolating low-bias Taylor expansions into the write regime.

Resonant double-barrier MTJs with embedded nanoparticles add another layer of nonlinearity. In the quantum-ballistic model FM/Insulator/NP/Insulator/FM, the nanoparticle acts as a quantum well with $64\%$1 and supports resonant transmission,

$64\%$2

This produces low-bias TMR suppression, narrow peak-like TMR anomalies of a few mV, and even simulated $64\%$3 for very small $64\%$4; the same structures can deliver much larger in-plane STT than a single barrier of the same total thickness (Useinov et al., 2016). In fully insulating MTJs, the dominant mechanism changes again: because transport is evanescent throughout the structure, the out-of-plane torque generally exceeds the in-plane torque, and both torques can increase by $64\%$5–$64\%$6 orders of magnitude at large bias due to spin-selective Fowler–Nordheim tunneling (Manchon, 2012).

4. Switching dynamics, write error, and assist mechanisms

The most direct figure of merit for memory use is the critical current or voltage needed to achieve low-error switching while preserving thermal stability and barrier integrity. In perpendicular Co-Fe-B/MgO MTJs, optimizing the out-of-plane bias field during DC-STT measurements reduced the average critical current density below $64\%$7 kA/cm$64\%$8, with a minimum of $64\%$9 kA/cm$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$00 at $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$01 mT. In the same device, the switching currents were $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$02 $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$03A for P$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$04AP and $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$05 $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$06A for AP$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$07P, while the extracted thermal stability factor was $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$08 for the low-$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$09 sample (Leutenantsmeyer et al., 2013). The study emphasized the familiar PMA trade-off: lower effective anisotropy reduces write current, but insufficient TMR and insufficient $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$10 compromise retention.

Write reliability at nanosecond timescales is strongly affected by nonuniform reversal. Micromagnetic simulations of $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$11 nm and $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$12 nm perpendicular MTJs with interfacial DMI showed that short-pulse write-error-rate curves can become non-monotonic. At $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$13 ns, $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$14 gives monotonic WER reduction with current, but $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$15–$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$16 mJ/m$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$17 produces a ballooning-like anomaly, and $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$18 mJ/m$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$19 keeps WER high across the tested current range because chiral multidomain textures persist after the pulse. Longer pulses, such as $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$20 ns, suppress the anomaly by allowing those textures to collapse (Das et al., 13 Nov 2025). This rules out a purely macrospin interpretation of short-pulse switching failure in such devices.

Several assist strategies aim to lower the DC write burden. One is thermal spin torque. In ultrathin-MgO junctions with a $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$21 nm barrier and RA about $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$22 $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$23m$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$24, transverse temperature gradients of about $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$25–$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$26 K/nm across MgO, corresponding to up to about $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$27 K across the barrier, shifted the AP$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$28P switching field by about $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$29–$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$30 Oe while leaving P$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$31AP essentially unchanged. Magneto-Seebeck measurements showed that the charge current associated with the temperature gradient would be about $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$32 too small to explain the effect through ordinary STT, linking the switching-field shift instead to a genuine thermal spin torque driven by conductance asymmetry near zero bias (Pushp et al., 2015).

Another route is radio-frequency preconditioning. In perpendicular MTJs with diameters $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$33, $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$34, $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$35, and $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$36 nm, applying a small RF pulse before the DC write pulse enhanced the switching probability relative to a DC-only baseline of $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$37. At $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$38 ns, the improvement $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$39 was $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$40, $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$41, $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$42, and $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$43 for those four diameters, and even at $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$44 the gains remained positive. Lower RF frequencies were more effective than excitation near the free-layer FMR, and the scheme allows shorter DC pulses without increasing the peak oxide stress when $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$45 (Hayward et al., 13 Dec 2025).

STT also remains functionally important in hybrid three-terminal MTJs. In time-resolved measurements on W/CoFeB/MgO devices, STT alone produced slower and less reproducible switching than SOT, but when combined with SOT and VCMA it helped accelerate domain-wall propagation, contributing to sub-nanosecond switching with a cumulative spread below $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$46 ns and standard deviation about $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$47 ns (Grimaldi et al., 2020). This does not replace two-terminal STT-MRAM physics, but it shows that STT retains a clear dynamical role in composite write schemes.

5. Oscillators and microwave functionality

STT-MTJs are also microwave sources, and in that context the angular structure of the torque is as important as its absolute magnitude. Ordered Mg-B-O interface engineering provides a route to strongly skewed torque, which Slonczewski-type analyses and the work of Rippard and co-workers associate with larger net energy input per precession cycle. In CoFe/Mg-B-O/CoFe junctions, ordered interface-doped barriers produced $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$48 values up to about $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$49, whereas thick undoped MgO or disordered B distribution restored near-symmetric torque (Tang et al., 2016). This is the device-level basis for using interfacial resonances to enhance STO output power.

A different oscillator architecture is the orthogonal nanocontact STNO based on an MTJ with a PMA CoFeB free layer and an in-plane reference layer. In that geometry, zero-field auto-oscillations occur at about $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$50 GHz, frequencies above $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$51 GHz are reached with applied field, and the frequency tunability can reach about $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$52 GHz/mA. The measured VCMA coefficient is about $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$53 fJ/(V·m), and the key conclusion is that $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$54 is governed mainly by VCMA rather than by STT, while damping-like STT mainly determines linewidth and power asymmetry (Jiang et al., 2019). A useful corrective to common intuition follows: in oscillator MTJs, STT is not necessarily the dominant determinant of frequency tunability even when it is indispensable for sustaining oscillation.

Capping-layer engineering alters this balance further. In the comparison between A-MTJs with free layer 2 CoFeB/0.21 Ta/6 CoFeSiB and B-MTJs with 2 CoFeB/0.21 Ta/7 NiFe, both families had comparable saturation magnetization and anisotropy field, but B-MTJs displayed lower damping and therefore lower auto-oscillation thresholds. A $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$55 nm B-MTJ nanopillar with a 10 Ta/7 Ru cap showed onset of auto-oscillation at about $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$56 mA, corresponding to $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$57 A/$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$58m$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$59, and emitted integrated microwave power in the microwatt range (Parvini et al., 2023). Ta capping also maximized TMR relative to Ru/Ru or Cu/Ru caps, linking microwave efficiency to the same interfacial engineering that governs memory readout.

6. Noncanonical STT-MTJs and expanded functionality

The STT-MTJ concept has broadened well beyond isolated binary memory cells. One extension is spin logic based on continuous free layers. Perpendicular MTJs interconnected through a single cross-shaped free layer were fabricated as a platform for spin-torque majority gates, with $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$60 nm MTJs separated by $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$61 nm and the free-layer edge extending $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$62 nm beyond each pillar. Independent $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$63 $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$64s voltage pulses produced P$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$65AP transitions in all four pillars, demonstrating local STT control of a shared magnetic network. Micromagnetic simulations then showed that if the cross is scaled so that its lateral size satisfies $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$66, where $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$67, majority-gate behavior becomes robust; with $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$68 A/m and $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$69 kJ/m$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$70, $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$71 nm and $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$72 nm (Wan et al., 2017). The present devices were limited mainly by W sidewall fencing and pinning, not by the logic principle itself.

Another extension is probabilistic computing. Superparamagnetic tunnel junctions with $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$73 nm in-plane free layers were measured with dwell times below $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$74 ns, a Poisson-fit average dwell time of $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$75 ns, and an autocorrelation time of $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$76 ns. In these devices, STT tunes the occupancy bias between 0 and 1, whereas Joule heating mainly sets the fluctuation rate; the switching rate rises substantially when the local temperature increases under current density of order $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$77 MA/cm$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$78 for RA about $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$79 $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$80m$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$81. Raw bitstreams fail standard randomness tests, but XOR$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$82 processing of four independent streams passes the NIST SP 800-22 suite (Schnitzspan et al., 2023). This makes explicit that in the superparamagnetic regime, STT is a control knob for stochasticity rather than a deterministic write mechanism.

At the opposite size extreme, atomic-scale resonant-tunneling MTJs realized in spin-polarized STM show that a single localized resonance can determine both spin filtering and STT. There the effective polarization

$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$83

controls the sign and magnitude of the torque. Experiments on individual Fe/W(110) and Co/Ir(111) nanomagnets found that changing either the bias or the injection position could reverse $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$84 and hence the STT without reversing the bias polarity (Bazarnik et al., 24 Oct 2025). This is direct microscopic evidence that resonant states can invert STT by energy selection alone.

A final misconception concerns antiferromagnets: that a spin-neutral tunneling current cannot drive a torque. PtMn$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$85MgO$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$86PtMn$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$87 all-antiferromagnetic tunnel junctions contradict that expectation. They exhibit room-temperature TMR of $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$88 and bidirectional switching at current densities of order $$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$89 MA/cm$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$90, explained by an imbalance between intra- and inter-sublattice spin currents or, equivalently, by the net cluster octupole polarization of each electrode (Kang et al., 3 Sep 2025). In ferrimagnetic Mn$$\mathrm{TMR}(V)=\frac{G_{\mathrm{P}}(V)-G_{\mathrm{AP}}(V)}{G_{\mathrm{AP}}(V)}.$$91Ga-based MTJs, first-principles calculations likewise revealed long-range oscillatory STT and, in the mirror-symmetric three-monolayer MgO case, resonant enhancement of both TMR and interfacial torque (Stamenova et al., 2020). Taken together, these results show that STT-MTJs are no longer adequately described as a single ferromagnet/MgO/ferromagnet technology class; they are a broader family of spin-transfer tunnel devices whose operative symmetry, resonance structure, and magnetic order can differ fundamentally.