Field-Free Current-Induced Switching of L1₀-$\mathrm{Fe}\mathrm{Pt}$ Using Interlayer Exchange Coupling for Neuromorphic Computing

$L{1}_0$-phase $\mathrm{Fe}\mathrm{Pt}$ is well known for its unusually robust perpendicular magnetic anisotropy (PMA) properties arising from strong conduction-electron spin-orbit coupling (SOC) with the $\mathrm{Fe}$ orbital moment. The strong PMA enables stable magnetic storage and memory devices with ultrahigh capacity. Meanwhile, SOC is also the premise of the recently discovered spin-orbit-torque (SOT) effect, which opens avenues for possible electrical manipulation of magnetization for $L{1}_0$-$\mathrm{Fe}\mathrm{Pt}$. The bulk SOT of the $L{1}_0$-$\mathrm{Fe}\mathrm{Pt}$ single layer was discovered recently; this leads to the magnetization of $L{1}_0$-$\mathrm{Fe}\mathrm{Pt}$ reversibly switching on itself. However, deterministic SOT switching of bulk perpendicularly magnetized $\mathrm{Fe}\mathrm{Pt}$ magnets relies on an external magnetic field to break the symmetry. Here, the symmetry-breaking issue is resolved by interlayer exchange coupling, where the $\mathrm{Fe}\mathrm{Pt}$ layer is coupled with an in-plane magnetized $\mathrm{Ni}\mathrm{Fe}$ layer through a $\mathrm{Ti}\mathrm{N}$ spacer layer. Furthermore, our device also presents memristive or gradual switching behaviors, even without an external field, offering the potential for constructing spin synapses and spin neurons for neuromorphic computing. An artificial neural network with high accuracy (∼91.17%) is realized based on the constructed synapses and neurons. Our work paves the way for field-free SOT switching of single bulk PMA magnets and their potential applications in neuromorphic computing.

Figure 1

$L{1}_0$-phase $\mathrm{Fe}\mathrm{Pt}$ and its current-induced switching. (a) XRD pattern of $\mathrm{Mg}\mathrm{O}$ (substrate)/$\mathrm{Fe}\mathrm{Pt}$ (gray line) and $\mathrm{Fe}\mathrm{Pt}/\mathrm{Ti}\mathrm{N}/\mathrm{Ni}\mathrm{Fe}$ heterostructure (dark-red line). Inset shows a schematic of the crystal structure of $L{1}_0$-$\mathrm{Fe}\mathrm{Pt}$. (b) Optical top-view image of the fabricated Hall bar device, and schematic of the current-induced-switching measurement setup. (c) AHE loop showing PMA of the $\mathrm{Fe}\mathrm{Pt}$ layer. Current-induced magnetization switching of $\mathrm{Fe}\mathrm{Pt}$ layer under different in-plane fields, which range from 200 to −200 Oe (d) and −200 to 200 Oe (e). Gray arrows indicate the sequence of applied in-plane fields. H_x and H_z represent the applied magnetic field along x and z axes, respectively.

Figure 2

Field-free switching with different IEC couplings. (a) HAADF image of the $\mathrm{Fe}\mathrm{Pt}(5)/\mathrm{Ti}\mathrm{N}(3.5)/\mathrm{Ni}\mathrm{Fe}(15)$ sample and its corresponding EDS mapping. Scale bar is 10 nm. (b),(c) Current-induced-switching process for the stack of $\mathrm{Fe}\mathrm{Pt}(5)/\mathrm{Ti}\mathrm{N}(2.3)/\mathrm{Ni}\mathrm{Fe}(15)$ with different sequences of applied H_x, which are indicated by gray arrows. Field-free switching is observed and the switching polarities at H_x = 0 Oe are opposite to that at H_x = 200 Oe in (b) or H_x = −200 Oe in (c), exhibiting AFM coupling. (d),(e) Current-induced-switching process for the stack of $\mathrm{Fe}\mathrm{Pt}(5)/\mathrm{Ti}\mathrm{N}(3.5)/\mathrm{Ni}\mathrm{Fe}(15)$, where FM couplings are observed. Note that (b)–(e) are the corrected switching loops after considering shunting effects (see Supplemental Material Note 1 [25]).

Figure 3

Evaluation of the spin-orbit effective fields. (a) Schematic of spin-orbit effective fields and the setup for harmonic measurement. (b) Field dependence of first- and second-harmonic signals for the $\mathrm{Fe}\mathrm{Pt}(5)/\mathrm{Ti}\mathrm{N}(2.3)/\mathrm{Ni}\mathrm{Fe}(15)$ heterostructure. Up and down triangular symbols correspond to +M_z and −M_z of the $\mathrm{Fe}\mathrm{Pt}$ layer, respectively. (c) Current-density dependence of longitudinal and transverse spin-torque effective fields of $\mathrm{Fe}\mathrm{Pt}(5)/\mathrm{Ti}\mathrm{N}(2.3)/\mathrm{Ni}\mathrm{Fe}(15)$ and $\mathrm{Fe}\mathrm{Pt}(5)/\mathrm{Ti}\mathrm{N}(3.5)/\mathrm{Ni}\mathrm{Fe}(15)$.

Figure 4

Memristive switching by electrical current pulses. (a) Cartoon of a biological synapse and neuron, and the measurement setup for memristive switching. (b) R_H-I loops with varying maximum current magnitude, I_max, under H_x = 0 Oe. Note that the applied maximum current value, namely, 46 mA, corresponds to the current density of 0.506 × 10⁷ A/cm². (c) R_H variations tuned by programming consecutive pulse sequences. Top panel illustrates the applied current pulses for tuning R_H, where equal numbers (80) of positive (+38 mA) and negative (−38 mA) pulses are alternately applied. Each pulse has a duration of 50 µs. Middle and bottom panels show R_H variations under current pulses with H_x = 200 and 0 Oe, respectively.

Figure 5

Sigmoidal artificial neuron. (a) Typical field-free switching loop, where data points in the shaded area are utilized for extracting the analog sigmoid function for an artificial neuron. (b) Data points selected from three repeated switching loops are used to obtain the analog sigmoid function. Black solid line represents the analog sigmoid function.

Figure 6

ANN based on the sigmoidal artificial neuron. (a) ANN consisting of three layers, where the input layer, hidden layer, and output layer have 784, 3, and 3 neurons, respectively. Sigmoidal artificial neuron extracted from experimental results serves as the hidden-layer neuron. 14 400 and 2400 images are used for training and testing, respectively. (b) Accuracy for training 14 400 images with different combinations of synapses and neurons.