Thermally nucleated magnetic reversal in CoFeB/MgO nanodots

Power consumption is the main limitation in the development of new high performance random access memory for portable electronic devices. Magnetic RAM (MRAM) with CoFeB/MgO based magnetic tunnel junctions (MTJs) is a promising candidate for reducing the power consumption given its non-volatile nature while achieving high performance. The dynamic properties and switching mechanisms of MTJs are critical to understanding device operation and to enable scaling of devices below 30 nm in diameter. Here we show that the magnetic reversal mechanism is incoherent and that the switching is thermally nucleated at device operating temperatures. Moreover, we find an intrinsic thermal switching field distribution arising on the sub-nanosecond time-scale even in the absence of size and anisotropy distributions or material defects. These features represent the characteristic signature of the dynamic properties in MTJs and give an intrinsic limit to reversal reliability in small magnetic nanodevices.

important role in determining the thermal stability of the system for in-plane dimensions below 50 nm, cannot be properly captured. The presence of interfaces causes the reduction of surface coordination and hence loss of exchange bonds at the surface, which leads to lower exchange coupling than in a bulk system. The micromagnetic approach tends to underestimate this effect and often only the dynamics of the free layer is considered. Determining the reversal mechanism is critical in evaluating the thermal stability and the switching time in spin transfer torque MRAM devices. Here we investigate the dynamics of CoFeB/MgO nanodots and MTJs at the atomistic level not constrained by the limits of micromagnetic models. The simulations demonstrate that the magnetisation reversal is incoherent for in-plane dimensions larger than 30 nm and that the switching of the magnetisation is driven by a thermal fluctuations. The atomistic thermal fluctuations at the edge of the device nucleate a domain wall which then propagates through the disk, leading to coercive fields significantly lower than for the case of coherent reversal. This new thermally induced reversal mechanism is a feature of atomistic simulations and are not seen shown by other micromagnetic computational methods. The fact that the switching is thermally driven poses an intrinsic limitation to the deterministic reversal process and reduces the thermal stability for small devices.

Results
Atomistic modelling of field induced magnetisation reversal. Using an atomistic spin model based on the Heisenberg Hamiltonian as implemented in the vampire software package 17,18 we simulate the dynamic properties of CoFeB/MgO nanodots, a schematic of which is presented in Fig. 1(a). Thermal effects are included in the model via a Gaussian white noise term whose amplitude is temperature dependent and the magnetostatic contribution using a modified macrocell approach. Details of the model are given in the methods section. We Figure 1. (a) Schematic of the simulated system with light spheres representing the high anisotropy layer, and dark spheres representing the bulk-like CoFeB layer. (b) Typical simulated easy-axis hysteresis loop for 1 nm thick, 50 nm diameter nanodot at temperatures of 5 K and 300 K. The data shows a large reduction in the coercivity for elevated temperatures due to increased thermal fluctuations, indicating a change in the magnetic reversal mechanism. (c) Snapshots of magnetisation reversal at 5 and 300 K for disk of diameter 50 nm and thickness 1 nm. I and IV refer to the top and bottom shoulder of M/M s vs H curve, respectively. II and III are configurations just before and after the switching, respectively. The colour scheme represents the magnetisation along the easy axis direction (z).
focus on the hysteretic properties of CoFeB/MgO nanodots, in particular investigating the temperature, size and thickness dependence of the reversal mechanism and dynamic coercivity. Finally, we consider the simulations of spin transfer torque switching in an MTJ device including the effects of atomistic thermal fluctuations.
The role of thermal fluctuations on the coercivity of CoFeB nanodots. We first consider the effects of temperature on the typical hysteresis properties of a nanodot with a diameter of 50 nm, as shown in Fig. 1(b). The first observation is that increased temperatures lead to a large reduction in the coercivity from 1.1 T at 5 K to 0.6 T at 300 K. The temperature variation of intrinsic properties such as the saturation magnetisation and magnetic anisotropy arises naturally from the atomistic simulations, using Monte Carlo methods as outlined in the methods section. This leads to an expected 20% reduction in H K between zero and 300 K but here we a observe a 45% reduction in the coercivity. This is partially due to the thermally activated transitions over the energy barrier, but also may reflect a change in the magnetic reversal mechanism due to the stronger thermal fluctuations. To investigate the reversal mechanisms we have generated snapshots of the atomic spin configuration during hysteresis for different temperatures, as shown in Fig. 1(c). At a temperature of 5 K the reversal is semi-coherent and nucleated at the centre of the nanodot due to the larger magnetostatic field. This is not generally observed by micromagnetic simulations and demonstrates the importance of the details of the magnetostatic coupling at these device dimensions. At 300 K the reversal is initiated by the nucleation of a small reversed domain at the edge of the nanodot caused by thermally driven spin fluctuations at the edge. At the edge a loss of exchange bonds leads to larger edge spin fluctuations compared with the spins in the middle of the dot. These larger spin fluctuations provide a natural nucleation region at the edges of the nanodot and therefore allow a different reversal mechanism compared to the centre nucleated reversal at low temperatures. Interestingly the small size of the system means that the thermal fluctuations are more important than the variation in the magnetostatic field across the dot diameter, highlighting the importance of including thermal fluctuations and surface effects in the model compared with non-stochastic continuum micromagnetic simulations. Due to the two different nucleation processes at low and high temperatures, the time that is required to reverse the magnetisation varies in the two temperature limits and the switching results faster at 5 K due to the semi-coherent nature of the mechanism.
We note that the thermally nucleated switching we describe here is different from the Sharrock approach 19 which considers a fixed (coherent) reversal mechanism but with a time dependence of the magnetisation due thermally induced transitions over the energy barrier. In the case of CoFeB/MgO dots the thermal fluctuations lead to a large reduction in the coercivity due to the ability to access a different thermally driven reversal mode. Of course, slower hysteresis loops will likely lead to a further reduction in the coercivity in a similar manner to that of Sharrock due to the increased number of nucleation attempts, but such simulations are currently beyond the time-scales accessible with atomistic models.
Another interesting feature of the hysteresis loop at 300 K in Fig. 1(b) is a slight asymmetry in the coercivity of the ascending and descending branches of the loop. This is due to the thermally nucleated nature of the reversal, leading to an uncertainty in the exact coercivity due to the randomness of the nucleation attempts. There is therefore an intrinsic thermal switching field distribution which is independent of defects and variations in the intrinsic properties, but arises solely due to random thermal fluctuations. For larger systems and long time-scales the thermal switching field distribution is not apparent, but for nanoscale MTJs switching in the nanosecond time domain it is a real and important effect and represents the thermodynamic limit of the switching field distribution which cannot be overcome.
Effect of nanodot diameter on the coercivity and thermal switching field distribution. To investigate the effects of nanodot size and temperature on the coercivity and thermal switching field distribution we have performed a systematic study of the hysteretic properties for 1 nm and 1.3 nm thick nanodots, shown in Fig. 2(a). The size dependence of the coercivity is obtained by averaging over a minimum of 30 independent loops for each size, temperature and thickness. The mean coercivity shows a complex temperature and size dependence which is due to different reversal mechanisms and finite effects. Considering first the 1 nm thick nanodots, the coercivity reaches an asymptotic limit for nanodot diameters >20 nm indicative of a nucleation reversal mode at 300 K with a slower approach at 5 K. However the snapshots of the atomic spin configurations support the earlier conclusion of different reversal modes at low and room temperature respectively. At 5 K the nucleation is driven by the variation of the magnetostatic field across the nanodot, which increases with increasing nanodot diameter leading to a slow convergence to a constant nucleation field only seen for larger nanodot diameters (around 100 nm, Supplementary Fig. 3, Supplementary Note 2). Conversely at 300 K the thermal nucleation volume is much smaller and independent of the dot size, and so the coercivity reaches an asymptotic limit at around 20 nm diameter. For dots smaller than 20 nm diameter the temperature has a dramatic effect on the coercivity, showing a large increase at 5 K and large decrease at 300 K respectively. We note that the increase of coercivity with decreasing diameter is indicative that the system has not reached the critical diameter for superparamagnetic behaviour.
For low temperatures the increase in the coercivity with decreasing diameter indicates a transition to coherent reversal (see Supplementary Fig. 2, Supplementary Note 1), where the magnetostatic field no longer dominates the reversal process and the nanodot size approaches the single domain limit δ π =Ã K / 1 0 w s eff nm. At room temperature the reduction in the coercivity is due to superparamagnetic fluctuations of the magnetisation which due to the small volume lead to switching at fields lower than the intrinsic coercivity. We note here that unlike the work of Brown 20 there is no peak in the coercivity due to the approximately two-dimensional nature of the nanodots and large anisotropy, hence the direct transition between superparamagnetic and nucleated reversal behaviour as a function of the nanodot size. For dots of diameter smaller than 10 nm the system enters in a single domain limit (in agreement with the estimation of the single domain size δ w ) and at room temperature the system becomes unstable due to a transition towards superparamagnetic regime. A similar size dependence is observed in other works 12,21 , although in the latter the investigated system has a lower effective anisotropy. This causes a larger critical single domain size and a less stable system, and therefore larger diameters are used. The 1.3 nm thick nanodots show a similar qualitative behaviour as the 1 nm thick nanodots as a function of the nanodot size, though with a significantly reduced coercivity. The large reduction in the coercivity arises mainly from the reduced anisotropy energy because of its proportionality with 1/thickness. In addition, a change in the magnetostatic energy caused by the increased thickness contributes to the decrease in the coercivity. The combination of these effects reduces the stability of the perpendicular orientation of the magnetisation and therefore increases the stability of nucleated domains under an applied field.
The statistical distribution of the coercivity for different nanodot sizes and temperatures is also strongly size dependent. The extracted switching field distributions (SFD) at room temperature for diameters of 10 and 50 nm and thickness 1 nm are presented in Fig. 2(b). The distributions show a range of switching fields which is much larger for the smaller nanodot size. In the case of our simulations, each nanodot of a given size is identical in terms of the number of atoms and magnetic parameters, but with different pseudorandom number sequence representing the random nature of the thermal noise in the simulations. Therefore, the origin of this distribution is purely the random thermal fluctuations during the reversal process, and hence the distribution is the thermal switching field distribution (TSFD) 22 . At the switching field the time scale of the reversal is determined by these random thermal fluctuations, leading to a natural TSFD for a switching process on the time-scale of a few nanoseconds. The TSFD is an intrinsic property of small magnetic elements and cannot be overcome due to its intrinsic thermodynamic origin. We note that the TSFD is also thickness dependent, being narrower for thicker films due to the reduced thermal fluctuations associated with the larger magnetisation volume. Importantly the TSFD intrinsically limits the ability to reliably reverse a nanodot at a given field and time-scale, leading to a natural distribution of switching probability for a finite time and strength of an applied field pulse 21 . For smaller diameters, the system becomes thermally unstable and the coercivity reduces at room temperature, while low temperature results in larger stability. (b) Calculated switching field distributions at 300 K for 10 and 50 nm nanodots. The data show that reduced nanodot diameters lead to a larger thermal switching field distribution affecting the stability of the magnetisation.  Fig. 3(a). Due to the strong coupling in MTJ, we have modified the usual macrocell approach for the calculation of the magnetostatic field following the approach proposed by Bowden 23 to obtain exact agreement with the atomic scale dipole-dipole interaction assuming a uniform magnetisation in each cell, a good approximation for our cell size of 1 nm 3 , as discussed in more detail in the methods section. We have calculated major and minor hysteresis loops for the MTJ structure at room temperature as shown in Fig. 3(b) and (c) respectively. We find that the free and pinned layers switch independently and that the reversal mechanism exhibits the same features observed for the individual layers, that of thermally nucleated switching (see Supplementary Figs 4 and 5, Supplementary Note 3). In major loops, compared to the single layer coercivities the magnetostatic coupling in the MTJ tends to stabilize the magnetic structure and enhances the coercivity of both layers compared to the free nanodots of about 0.1 T and 0.2 T for pinned and free layer, respectively. In the minor loop, shown in Fig. 3(c), the free layer exhibits a bias due to the stabilising (destabilising) effect of the magnetostatic field from the pinned layer for the descending (ascending) branches. To quantify the the magnetostatic field from the pinned layer acting on the free layer we have calculated the stray field with atomic resolution as function of position and the net average stray field in Fig. 4, showing the existence of a stabilising (destabilising) field depending on magnetic configuration.   In the case of MTJ devices the strong coupling of the magnetic layers leads to a a complex change in the magnetic properties such as coercivity. We have also investigated the effect of thermal fluctuations on spin transfer torque switching mechanism following Slonczewski's approach 24 . Figure 5 shows the time evolution of magnetisation for a MTJ of diameter 40 nm at room temperature. We observe that the magnetisation is reversed in the order of a nanosecond, in agreement with switching times measured experimentally by Devolder et al. 25 and Hahn et al. 26 . From the analysis of the spin configurations during the spin transfer torque switching (see Supplementary  Movie 1, Supplementary Note 4), a thermally activated incoherent reversal occurring via domain nucleation at the edge of the dot emerges. The edge nucleated nature of the reversal agrees with the reversal mechanism induced by an applied field as discussed previously. A stochastic nature of the magnetisation reversal is also found in the works of Devolder et al. 27 and Hahn et al. 26 . There they investigate the magnetisation switching due to spin transfer torque by means of time-resolved measurements and find that the spin transfer torque switching is thermally activated for comparable in-plane dimensions of MTJs. On the other hand, it is not possible to access the reversal mechanism of the magnetisation experimentally for such time-scales and therefore conclusions relative to the switching mechanism are possible only via indirect observations. Devolder et al. 27 find that for 100 nm and smaller MTJs the switching is irreversible with a weak dependence of the switching time on the device area and they explain it assuming nucleation of a domain at the edge of the system which then sweeps through the device. Similar analysis and conclusion are presented by Hahn et al. 26 . It is worth pointing out that Hahn et al. expect a change in the switching mechanism for diameters of about 50 nm, close to the estimate of the critical diameter for their system. Given our set of parameters, we expect this to occur at smaller dimensions, between 10 and 20 nm, as also shown in the case of hysteresis loops. Therefore, experiments performed on spin transfer torque dynamics for similar MTJ stacks of comparable dimension confirm the stochastic nature of the switching excluding a macro-spin nature of the reversal but rather a non-collinear mechanism characterised by domain wall nucleation and propagation. Nonetheless, we stress that the reversal mechanism we observe is characterised by a precessional motion where both the x and y-components oscillate in time and not by a pure domain sweeping through the system, as it occurs in hysteresis simulations and is assumed experimentally. Hence, atomistic simulations provide an insight into the nature of the reversal mechanism that would not be accessible otherwise experimentally due to the fast time-scale at which the spin torque dynamics occurs and by micromagnetic models because limited to 0 K. Interestingly, no large differences in the mechanism of the switching of the magnetisation at low and high temperature are found. This seems to suggest that the spin torque field acts favouring nucleation modes at the edge of the system and therefore edge nucleation should be expected.

Discussion and Conclusion
We have investigated the magnetisation reversal mechanism in CoFeB/MgO nanodots and MTJs using an atomistic spin model with the inclusion of thermal and magnetostatic effects. The magnetisation reversal in CoFeB/ MgO nanodots and MTJs can be described as thermally nucleated and incoherent at temperatures relevant to device operation, leading to a large reduction in the coercivity compared to a coherent reversal mechanism. The thermal nature of the reversal mechanism is also reflected in the spin transfer torque switching mechanism of MTJ devices, hence affecting the reversal speed. In an MTJ geometry we find that the magnetostatic interaction between the layers leads to a stabilising effect on both the pinned and free layers and causes a shift of the minor hysteresis loop. Our results highlight the importance of considering finite size and thermal fluctuations when modelling such small scale magnetic devices which can have a dominant effect on their reversal mechanisms and physical properties. It is important to note the large difference between the coercivities in our simulation of a perfect nanodot and those measured experimentally, where coercivities are typically ~0.1 T. In our model we have used material parameters derived from experimental measurements of continuous thin films, and as such our simulations represent the best case situation concerning large coercivity. We expect that realistic devices are affected by edge damage and defects which lead to a further reduction of the coercivity. Our results also raise further questions on the role of thermal fluctuations on spin transfer torque switching and the energy barrier in zero field responsible for the thermal stability of MTJs. We expect that the time-scale of the spin transfer torque switching is strongly dependent on the lateral size of the MTJ due to different magnitudes of the thermal fluctuations breaking the magnetic symmetry required for switching and will be the subject of future work.

Methods
Atomistic spin model. The simulations were performed using an atomistic spin model where the energy of the system is described by a classical Heisenberg Hamiltonian ( ) where S i,j are normalized spin vectors on site i, j respectively, J ij is the exchange coupling between spin i and j, k u is the single-ion uniaxial magnetocrystalline anisotropy energy (MAE) constant per site, μ s the atomic spin moment, H app the applied external field and  demag the magnetostatic contribution. First and second term on RHS of equation 1 describe a system with nearest neighbours isotropic exchange interactions and uniaxial MAE respectively, while the third term represents the Zeeman interaction with an external field 18 . Given the high computational cost required to calculate calculate the magnetostatic energy due to the long range nature of this interaction and because the fluctuation of the exchange energy are generally larger than that of the magnetostatic energy, the demagnetisation field is computed applying a micromagnetic discretisation of the system into macrocells that are considered as dipoles. Each macrocell i has a magnetic moment m i mc determined by the vector sum of the atomic spin moments inside the cell and position calculated from the magnetic centre of mass of the cell  33 . The value of the atomic spin moment in our simulations is significantly lower than expected experimentally for bulk CoFe or from ab-initio calculations of CoFe/MgO, where values close to 2.5 μ B are found 34,35 . Experimentally the CoFeB/MgO system is known to have a perpendicular orientation for effective thicknesses less than 1.2-1.3 nm 6,11,33,[36][37][38][39][40] and hysteresis simulations for different atomic moments ( Supplementary Fig. 1, Supplementary Methods) confirm that an effective atomic moment less than 2 μ B is required to have perpendicular orientation of the magnetisation and square loops. The physical origin of the reduced saturation magnetisation is likely due to a combination of the presence of non-magnetic Boron and the possibility of structural defects in the material. The effect of the demagnetising field is included in the calculations using a macrocell approach 17 with a cell size of 1 nm. The used parameters are reported in Table 1.
In the hysteresis loop calculations we use a critical damping and calculate a complete hysteresis cycle over 20 ns with an effective field rate of 0.3 T ns −1 to minimize the effects of enhanced coercivity caused by fast field sweep rates.