France



Two-dimensional magnetic recording (TDMR) with laterally-displaced multiple read heads has been developed to alleviate the impact of inter-track interference (ITI) and track edge effects. However, the cross-track (XT) position of the displaced 2&lt;sup&gt;nd &lt;/sup&gt;head varies significantly with the skew angle of head assembly from ID to OD across disk radius [1]. As a result, each disk platter is divided into hundreds of groups of sectors, or zones, to accommodate the skew angle variations. The 2-dimentional equalization technique used in conventional data detection channels requires different channel parameters at different skew angles/zones for channel optimization, and the number of different channel parameter sets across the disk platter are in hundreds. Furthermore, resulting performance gain is further diminished by random fluctuations such as non-repeatable disk runout.&lt;br/&gt;&lt;br/&gt;In this situation, if we employ machine learning technique on TDMR data detection, one could imagine the possibility of achieving optimal performance in resolving ITI and track edge effects across the entire disk surface while using “learning” to replace relatively complex parameterization of physical modeling. In our previous work, we have developed deep neural network (DNN) for disk drive data detection channels with performance (detection error-rate vs. signal-to-noise ratio) slightly more superior than the state-of-the-art disk drive channels [2,3]. In the work presented here, we train a convolution neural network (CNN) in TDMR system to detect data under a wide range of ITI from different head skew angles. We demonstrate that in the medium noise dominant situation, CNN can learn all ITI in the training dataset and completely recover ITI-free BER without any physical modeling. Our eventual goal is to explore if we can train a single CNN detector to learn and optimize for all head skew angles/zones automatically, thereby lifting the complexity of designing different channel parameter sets at each head skew angle.&lt;br/&gt;&lt;br/&gt;The recorded tracks and CNN detection scheme are shown in Fig. 1. Two adjacent data tracks are written with random data and arbitrary phase. The width of the track is denoted &lt;i&gt;TW&lt;/i&gt;, and no erase-band nor transition curvature is included. Two XT overlapped heads of width greater than &lt;i&gt;TW&lt;/i&gt; are used. Here we simplify the modeling and assume each head only captures signals under its XT coverage. Since the relative positions of head-1 and head-2 are locked, we characterize ITI of the reader array with a single variable: the ratio between head-1 coverage of Track-II (&lt;i&gt;Δx&lt;sub&gt;1&lt;/sub&gt;&lt;/i&gt;) versus the track width (&lt;i&gt;TW&lt;/i&gt;): &lt;i&gt;δ&lt;sub&gt;1&lt;/sub&gt;=Δx&lt;sub&gt;1&lt;/sub&gt;/TW&lt;/i&gt;. Here we assume the ITI stays fixed within each zone, and different zones will have different ITI. In this work, we only investigate the medium noise dominated situation: the jitter noise is modeled as the uncertainty of the transition locations following a Gaussian distribution. Since both heads “see” the same magnetic patterns from the two tracks, the jitter noise is correlated noise. A CNN is used as the data detector in this work. It receives signals from both heads as input, and we only train it to detect data for Track-I. The rest of the parameters are listed in Fig. 1 caption.&lt;br/&gt;&lt;br/&gt;To simulate reading data at different skew angles, we evaluate the CNN detection BER with signals of 11 different ITI from &lt;i&gt;δ&lt;sub&gt;1&lt;/sub&gt;&lt;/i&gt;=0.2 to 0.8 as in Fig. 2. The black dashed line benchmarks the ITI-free BER. For the blue curve, the CNN detector is trained using signals from one skew angle/zone with a fixed ITI, &lt;i&gt;δ&lt;sub&gt;1&lt;/sub&gt;&lt;/i&gt;=0.5 . In this case, CNN can only eliminate the trained ITI, and the detection accuracy degrades significantly whenever the testing ITI is different. In contrast, the red curve plots the BER when we train the CNN using signals with mixed ITI same as the testing dataset. This time CNN can successfully eliminate ITI for all testing head skew angles/zones.&lt;br/&gt;&lt;br/&gt;Our results suggest that CNN has the capacity to learn very complicated correlations and is solely data-driven. With the learning process, the CNN-based ML channel presented here is capable of resolving the interference from ITI infested signals, automatically rejecting them, and arriving at correct detection. This approach is fundamentally different from the conventional channel design that relies on sophisticated physical modeling and channel parametrization. In addition, the fact that a single CNN can learn a wide range of ITI states that it can optimize the channel performance for head skew angles using a single CNN eliminating the need to use multiple channel parameter sets. Not only this machine learning channel can eliminate the complex physical characterization process required for conventional approach, more importantly, the TDMR performance gain can be fully realized without compromising. The potential of fully eliminating ITI also presents the possibility to have reader width wider than track width which could be critically important as track pitches in HDDs continue to decrease.

**References:**

[1] Suzuto R, Nakamura Y, Osawa H, et al. Effect of reader sensitivity rotation in TDMR with head skew. IEEE Transactions on Magnetics, 2016, 52(7): 1-4.&lt;br/&gt;[2] Qin Y, Zhu J G. Deep Neural Network: Data Detection Channel for Hard Disk Drives by Learning. IEEE Transactions on Magnetics, 2020, 56(2): 1-8.&lt;br/&gt;[3] Qin Y, Zhu J G. Automatically Resolving Inter-Track Interference with Convolution Neural Network Detection Channel in TDMR. IEEE Transactions on Magnetics, 2020.

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/1636e831501326557b33b2a856085d21.jpg)
CNN detector in TDMR. The heads have DT sensitivity of PW50/bit-length = 2.5. The green(yellow) signals are from head-1(head-2), and they are sampled 1 bit/period. The jitter width is assumed 1.4nm in this work. The CNN channel has 3 hidden layers (blue) with 50 filters each. The kernel sizes are 3-3-3-1. &lt;i&gt;ReLU&lt;/i&gt; is used as the non-linear activation function, and cross-entropy is the loss function.

![](https://assets.underline.io/uploads/markdown_image/1/image/3a69ad272e492b195d37c70a85e1048b.jpg)
Detection BER at different zone/ITI. Black dashed curve: ITI-free detection BER. Blue: CNN is only trained with signals of fixed ITI &lt;i&gt;δ&lt;sub&gt;1&lt;/sub&gt;=0.5&lt;/i&gt; . Red: CNN is trained with signals with 11 different ITI (same as the testing dataset).

INTERMAG 2021

 Enable TDMR Gain with Convolution Neural Network Based Machine Learning Algorithm

INTERMAG 2021 Welcome Video



	
<figure class="video_container">
 <iframe src="https://screencast-o-matic.com/player/crflqFVnUpZ?v=6&ff=1&title=0&controls=1" width=640 height=360 frameborder="0" allowfullscreen="true"> </iframe>
</figure>

Want to buy a ticket? 

Register for INTERMAG 2021 here

[![](https://assets.underline.io/uploads/markdown_image/1/image/17ce76e7960db633258024eef8d6da7f.png)](https://intermag.org/register)

Got your ticket, but still can’t access the event?

Contact us and our team will look into it.

Two-dimensional magnetic recording (TDMR) with laterally-displaced multiple read heads has been developed to alleviate the impact of inter-track interference (ITI) and track edge effects. However, the cross-track (XT) position of the displaced 2nd head varies significantly with the skew angle of head assembly from ID to OD across disk radius [1]. As a result, each disk platter is divided into hundreds of groups of sectors, or zones, to accommodate the skew angle variations. The 2-dimentional equalization technique used in conventional data detection channels requires different channel parameters at different skew angles/zones for channel optimization, and the number of different channel parameter sets across the disk platter are in hundreds. Furthermore, resulting performance gain is further diminished by random fluctuations such as non-repeatable disk runout. In this situation, if we employ machine learning technique on TDMR data detection, one could imagine the possibility of achieving optimal performance in resolving ITI and track edge effects across the entire disk surface while using “learning” to replace relatively complex parameterization of physical modeling. In our previous work, we have developed deep neural network (DNN) for disk drive data detection channels with performance (detection error-rate vs. signal-to-noise ratio) slightly more superior than the state-of-the-art disk drive channels [2,3]. In the work presented here, we train a convolution neural network (CNN) in TDMR system to detect data under a wide range of ITI from different head skew angles. We demonstrate that in the medium noise dominant situation, CNN can learn all ITI in the training dataset and completely recover ITI-free BER without any physical modeling. Our eventual goal is to explore if we can train a single CNN detector to learn and optimize for all head skew angles/zones automatically, thereby lifting the complexity of designing different channel parameter sets at each head skew angle. The recorded tracks and CNN detection scheme are shown in Fig. 1. Two adjacent data tracks are written with random data and arbitrary phase. The width of the track is denoted TW, and no erase-band nor transition curvature is included. Two XT overlapped heads of width greater than TW are used. Here we simplify the modeling and assume each head only captures signals under its XT coverage. Since the relative positions of head-1 and head-2 are locked, we characterize ITI of the reader array with a single variable: the ratio between head-1 coverage of Track-II (Δx1) versus the track width (TW): δ1=Δx1/TW. Here we assume the ITI stays fixed within each zone, and different zones will have different ITI. In this work, we only investigate the medium noise dominated situation: the jitter noise is modeled as the uncertainty of the transition locations following a Gaussian distribution. Since both heads “see” the same magnetic patterns from the two tracks, the jitter noise is correlated noise. A CNN is used as the data detector in this work. It receives signals from both heads as input, and we only train it to detect data for Track-I. The rest of the parameters are listed in Fig. 1 caption. To simulate reading data at different skew angles, we evaluate the CNN detection BER with signals of 11 different ITI from δ1=0.2 to 0.8 as in Fig. 2. The black dashed line benchmarks the ITI-free BER. For the blue curve, the CNN detector is trained using signals from one skew angle/zone with a fixed ITI, δ1=0.5 . In this case, CNN can only eliminate the trained ITI, and the detection accuracy degrades significantly whenever the testing ITI is different. In contrast, the red curve plots the BER when we train the CNN using signals with mixed ITI same as the testing dataset. This time CNN can successfully eliminate ITI for all testing head skew angles/zones. Our results suggest that CNN has the capacity to learn very complicated correlations and is solely data-driven. With the learning process, the CNN-based ML channel presented here is capable of resolving the interference from ITI infested signals, automatically rejecting them, and arriving at correct detection. This approach is fundamentally different from the conventional channel design that relies on sophisticated physical modeling and channel parametrization. In addition, the fact that a single CNN can learn a wide range of ITI states that it can optimize the channel performance for head skew angles using a single CNN eliminating the need to use multiple channel parameter sets. Not only this machine learning channel can eliminate the complex physical characterization process required for conventional approach, more importantly, the TDMR performance gain can be fully realized without compromising. The potential of fully eliminating ITI also presents the possibility to have reader width wider than track width which could be critically important as track pitches in HDDs continue to decrease.

**References:**

[1] Suzuto R, Nakamura Y, Osawa H, et al. Effect of reader sensitivity rotation in TDMR with head skew. IEEE Transactions on Magnetics, 2016, 52(7): 1-4. [2] Qin Y, Zhu J G. Deep Neural Network: Data Detection Channel for Hard Disk Drives by Learning. IEEE Transactions on Magnetics, 2020, 56(2): 1-8. [3] Qin Y, Zhu J G. Automatically Resolving Inter-Track Interference with Convolution Neural Network Detection Channel in TDMR. IEEE Transactions on Magnetics, 2020.

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/1636e831501326557b33b2a856085d21.jpg)
CNN detector in TDMR. The heads have DT sensitivity of PW50/bit-length = 2.5. The green(yellow) signals are from head-1(head-2), and they are sampled 1 bit/period. The jitter width is assumed 1.4nm in this work. The CNN channel has 3 hidden layers (blue) with 50 filters each. The kernel sizes are 3-3-3-1. ReLU is used as the non-linear activation function, and cross-entropy is the loss function.

![](https://assets.underline.io/uploads/markdown_image/1/image/3a69ad272e492b195d37c70a85e1048b.jpg)
Detection BER at different zone/ITI. Black dashed curve: ITI-free detection BER. Blue: CNN is only trained with signals of fixed ITI δ1=0.5 . Red: CNN is trained with signals with 11 different ITI (same as the testing dataset).

technical paper

Magnetic skyrmion is a non-trivial topological object which originates due to the interplay among various types of interactions such as magneto-static, anisotropy, dipolar, exchange interaction, and Dzyaloshinskii-Moriya interaction (DMI). In magnetic thin films, skyrmions originate through the competition between interfacial DMI and exchange interaction1. Patterned nanostructures provides a pathway to predictably stabilize and thus allow understanding their behavior in presence of different important parameters in a more controlled way (ref.). Moreover, due to the potential application of skyrmions in spintronic devices such as race-track memory etc., several studies of their creation, stability, and dynamics in nano wires have been reported1. Skyrmions in nanostructures can be dynamically controlled by magnetic field gradiant3 and spin-polarized current1. In case of high-speed skyrmion-based magnetic storage devices, large spin-polarized current density is required to control skyrmion dynamics. This in turn causes Joule heating and skyrmion Hall effect4 which limits the efficiency of skyrmion-based devices. These issues can be resolved by employing magnon spin current as a tool to control skyrmion propagation. However, very few studies have so far reported dynamics of skyrmions in presence of temperature gradient in a nanotrack5. L. Kong et al. reported the systematic investigation of the dynamics of Bloch skyrmions in presence of the temperature gradient. Skyrmions are found to move from region of lower to higher temperature. Moreover, the velocity of skyrmions is found to increase with the increase of thermal gradient5. Magnon spin current generated due to thermal gradient diffuses from hot to the cold region. This magnon spin current applies a negative torque on the skyrmion resulting in movement of skyrmion cold to hot region4. This study also reports that the skyrmion Hall component is absent or very small with the magnon spin current. A few other studies report the role of temperature on the current induced dynamics which demonstrate the velocity of skyrmion for the constant current density increases with the temperature 7. Although the dynamical behavior of Bloch skyrmions in presence of a temperature gradient has been reported in literature, there is no comprehensive studies performed to understand the dynamics of Neel skyrmion in presence of temperature gradient. In this work, we performed micromagnetic simulations to investigate the dynamics of Néel skyrmions in a race track of Co/Pt in presence of the temperature gradient. We considered a Co/Pt track with dimensions of 100 nm x40 nm x 1nm. Individual skyrmions were stabilized due to the i-DMI of ~3 mJ/m2 in the nanotrack and thereafter, their dynamics in presence of a temperature gradient7,8 was studied. Our studies reveal that similar to Bloch skyrmions, the temperature gradient plays a significant role on the velocity of Neel skyrmions in the nanotrack. The skyrmions move from cold to the hotter region and the velocity increases as the temperature gradient is increased (Fig. 1(b)). Furthermore, we have also investigated the role of Gilbert damping constant on the skyrmion velocity in presence of thermal gradient on the track. Our results show that the velocity is inversely proportional to the values damping constant. The details of these results are discussed in this work.

**References:**

1 J. Iwasaki, et al. Nat. Nanotechnol. 8, 742 (2013). 2 A. Casiraghi, et al. Nat. Commun Phys. 2, 145 (2019). 3. J. J. Liang et al. New J. Phys. 20, 053037 (2018). 4 G. Chen, et al. Nat. Phys 13, 112–113 (2017). 5 L. Kong et al. Phys. Rev. Lett. 111, 067203, (2013). 6 M T Islam et al. J. Phys.: Condens. Matter 31, 455701 (2019). 7 Litzius, K. et al. Nat Electron 3, 30–36 (2020). 8 Everschor, K. et al. Phys. Rev. B 86, 054432 (2012).

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/33fa4a9ffb9f74e344b2669dae0c9fd4.jpg)
Fig.1.(a) Position of a skyrmion in a nanotrack in presence of temperature gradient at three different times. (b) The velocity of Skyrmion as a function of the temperature gradient in K/nm.

 Thermal gradient driven dynamics of Neel skyrmions in a nanoracetrack

Skyrmions have been proposed for racetrack memory [1], but there are still many questions about the best way to detect them electronically. The most promising approaches so far have used the transverse voltage due to the skyrmion Hall effect, but the signal is still very weak [2, 3]. Here we use simulations consider alternative vertical detection methods using magnetoresistance. A spin polarized current is used to move a Néel skyrmion down a racetrack that contains a circular magnetic dot above it. When the dot is flush with the racetrack, exchange interactions dominate, while when it is separated by a 1 nm thick tunnel barrier, the interactions are mainly magnetostatic. The trajectory of the skyrmion over a period of ~ 4 ns is calculated using mumax3micromagnetics simulation software. Results for different dot and skyrmion core magnetization directions are used as a guide to predict the magnetoresistance as a function of time, which would be used to detect individual skyrmions. For these model studies, a saturation magnetization of 5.8 x 105 A/m, exchange stiffness 1.5 x 10-12 J/m, anisotropy 8 x 105 J/m3, and damping parameter 0.1 were used for both the racetrack and 20 nm diameter detection dot. In addition, to stabilize 20 nm diameter skyrmions, the racetrack also had a Dzyaloshinskii-Moriya Interaction strength of 3 x 10-3 J/m2. A current density of 1 x 1012 A/m2 with spin polarization 0.4 was used to move the skyrmions. In all cases there was a transverse deflection due to the skyrmion Hall effect. When the dot is flush with the racetrack, exchange coupling to the skyrmion spin texture background can be treated as an effective magnetic field with an antiskyrmion texture [4]. When there is a 1 nm MgO tunnel barrier separating the dot and racetrack, magnetostatic effects dominate and the magnitude of the interaction field is much smaller. However, simulations show that the vertical tunnel current is dominated by the volume directly below the dot. There will be a transient change in the magnetoresistance as a skyrmion passes by the dot detector. With magnetostatic coupling both the trajectory and skyrmion size as a function of time differ, depending on whether the dot and skyrmion core magnetizations are parallel or antiparallel. When they are parallel (Figure 1), the skyrmion diameter grows slightly as it approaches the dot, since the dipolar field tends to orient misaligned spins. The spatial field gradient due to the dot causes a repulsive deflection around the dot. When the skyrmion core and dot magnetiation are antiparallel (Figure 2), the skyrmion slightly shrinks in diameter as it approaches the dot, and there is an attractive interaction. The magnitudes of the deflection depend on the dot and skyrmion stack diameters and thicknesses, but as long as part of the skyrmion passes under the dot, it should also be possible to differentiate core up and core down skyrmions. For Figure 1 there would be a transient drop in resistance, and for Figure 2 a transient rise. If in-plane magnetization dot detectors were used, there would also be characteristic magnetoresistance transients that are sensitive to the in-plane component of the skyrmion spins. This could enable Néel and Bloch skyrmions to be differentiated, if magnetized along the racetrack axis, and chirality detection of Bloch skyrmions, if magnetized perpendicular to the racetrack axis.

**References:**

1. A. Fert, V. Cros, J. Sampaio, Nat. Nanotech. 8 152 (2013). 2. D. Macciarello, et al., Nat. Nanotech. 13, 233 (2018). 3. K. Zeissler, et al., Nat. Commun, 11, 428 (2020). 4. K. Hamamoto, M. Ezawa, and N. Nagaosa, Phys. Rev. B 92, 115417 (2015).

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/169f90678345b92ecec0d1db3eed2d01.jpg)
20 nm Néel skyrmion trajectory along a 200 nm x 80 nm racetrack, passing by a 20 nm diameter dot separated by a 1 nm tunnel junction barrier. The skyrmion core and dot have parallel magnetizations. a) at beginning, b) repulsion near the dot, and c) at the end of the racetrack.

![](https://assets.underline.io/uploads/markdown_image/1/image/1fd3076c0db495724367847365e0bfa1.jpg)
Similar conditions to those of Figure 1 except that the dot magnetization direction has been reversed, so that it is antiparallel to that of the skyrmion core. Here the interaction near the dot is attractive.

 Simulations of Magnetoresistive Detection of Skyrmions

Second harmonic Hall voltage (SHV) method [1] has been widely used to measure spin-orbit torques (SOTs) in heavy metal/ferromagnet (HM/FM) bilayers. It is typically applied for the single domain regime while the signals in multi-domain regimes are almost always ignored due to the complication from domain structures. In this study, we analyze the SHV data taken in a magnetic multilayer made of W/Pt/Co/SiO2. First harmonic planar Hall effect (PHE) and anomalous Hall effect (AHE) measurements reveal weak in-plane anisotropy of the Co layer. At low temperatures, the PHE loops become asymmetric, which indicate sizable Dzyaloshinskii-Moriya interaction (DMI) at the Pt/Co interface due to inversion symmetry breaking and strong spin-orbit coupling [2]. Typical sheared hysteresis for labyrinth domains is observed in AHE measurements. Interestingly, the second harmonic Hall voltage loops show “humps” in the intermediate field range where DMI-induced topological Hall effect and magnetic bubbles are typically observed. We explain such enhancement in SHV as a result of chiral Néel wall formation and efficient SOT driven domain wall motion [3]. Our results demonstrate that second harmonic voltage measurement can be utilized to detect chiral spin structures in magnetic multilayers.

**References:**

[1] M. Hayashi et al, Phys. Rev. B 89, 144425 (2014). [2] M. Heide et al, Phys. Rev. B 78, 140403(R) (2008). [3] A. V. Khvalkovskiy et al, Phys. Rev. B 87, 020402(R) (2013).

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/d4f5c98afb392aa4644f93209f07ec09.jpg)
Fig.1 Planar Hall effect measurement under in-plane field results for a W(3)/Pt(1)/Co(1.2)/SiO2(5) device.

![](https://assets.underline.io/uploads/markdown_image/1/image/69bc664348995b3c5e32aa62bbd1c7ca.jpg)
Fig.2 First and second harmonic Hall voltages under a nearly out-of-plane field scan. FM, Néel, and HTH denote ferromagnet, Néel domain wall, and head-to-head domain wall, respectively.

 Second harmonic detection of chiral spin structures in a magnetic multilayer

In the chiral ferromagnets with the Dzyaloshinskii-Moriya (DM) spin-orbit interaction, vortex spin microstructures arise, which are considered as a unit of information for computer memory. In strongly correlated itinerant ferromagnets, these spin structures arise due to the first-order thermodynamic phase transitions, in which an intermediate phase of the chiral spin liquid is realized. However, the experiments indicate the realization of a more complex and still unexplained picture of spin microstructures, which depends not only on temperature and external magnetic field, but also on pressure. In this communication, we investigate the conditions for the occurrence of such spin microstructures in the framework of the spin-fluctuation theory of strongly correlated band ferromagnets with the DM-interaction. We consider the Hubbard model and DM-interaction, which, due to the small relativism, is considered in the mean-field approximation. The resulting expression for the free energy functional contains pressure-dependent magnetic, electronic, and spin-fluctuation terms, which are self-consistent with each other. The conditions for the minimum of free energy and electroneutrality are considered at various pressures in the model of the electronic structure of MnSi obtained from the LDA+U+SO calculations. The effect of pressure on the DOS determined by LDA+U+SO-calculation will be considered in the model of corresponding states. The initial state is chiral ferromagnetic helicoidal spin spiral (CFH) phase. An increase in temperature and pressure leads to thermodynamic and quantum transitions, which are responsible for the formation of the MnSi phase diagram (see fig.). A crossover of a first-order thermodynamic and quantum phase transition is obtained, which leads to the appearance of a chiral spin fluid (CSF). An increase in pressure suppresses local DOS minimum and zero-point spin fluctuations. This leads to the disappearance of the crossover of quantum and thermodynamic phase transitions at the multicritical boundary point (T0, p0). Below the temperature T0, at p=p0, a quantum transition to the state of phase segregation (SP) occurs. This transition is accompanied by an increase in the DOS at the Fermi level and leads to the appearance of a ferromagnetic state. In this state, due to the DM-interaction and fluctuating exchange fields, skyrmions of the left and right spin chirality appear. Moreover, in the SP-region baric magneto-volumetric effect is realized, which is observed experimentally. An increase in pressure at T&lt;T0 and p&gt;p0 leads to the crossing of the DOS maximum by the chemical potential and a quantum transition to a paramagnetic state with a nonzero local magnetization (CPM). In the CPM-region (p&gt;pS(T)), the spin chirality fluctuations are realized. An increase in the amplitude of spin fluctuations with increasing pressure in this phase ensures the disappearance of local magnetization. As a result, at p=pC(T), a second-order phase transition to the paramagnetic phase (PM) without spin chirality occurs. The (p-T)-diagram of MnSi, calculated in the framework of the developed spin-fluctuation approach using the LDA+U+SO spectrum, is consistent with experiment.

**References:**



**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/a75fc5d7b43c1a1ef3f59af9aaec7ab5.jpg)

 Quantum phase transitions under pressure in a chiral itinerant ferromagnet MnSi

Microwave-assisted magnetic recording (MAMR) is a new technology where a microwave magnetic field having a frequency of few tens of GHz can assist switching of the magnetic domains in the recording media[1]. This microwave magnetic field is generated by using a spin-torque oscillator (STO) under the influence of the magnetic field from the write pole of the recording head and an applied dc bias current. When certain conditions for bias current and magnetic field are met, spin-transfer torque causes the field generating layer (FGL) of the STO to precess, generating an oscillating magnetic field. Although the oscillation frequency of the magnetic field generated by the STO is a key parameter in MAMR, its conventional measurement using the radio-frequency (RF) electrical signal arising from the magnetoresistance (MR) of STO has the following concerns: the RF electrical signal is sometimes very weak and contains several frequency components that do not correspond to the frequency of the actual magnetic field generated by the STO because of the complex magnetization dynamics of the ferromagnetic layers inside STO; the RF signal attenuates heavily in the write head-STO assembly. Therefore, it is strongly desired to find a way to characterize the oscillation frequency of the STO which is already a part of the write head assembly. In this study, we have proposed and successfully demonstrated a new technique to characterize the actual frequency of the magnetization precession in an STO using injection locking from externally applied radio-frequency (RF) magnetic field. Using our method, there is no need to electrically connect the STO to a spectrum analyzer via an RF-compatible electrical circuit. Injection locking is a phenomenon occurring when an oscillator (in this case, an STO) is forced to couple and synchronize with an external oscillating signal (in this case, oscillating magnetic field at GHz frequency range)[2]. When the FGL of an STO is precessing at a given frequency (fFGL), it can undergo injection locking to an externally applied magnetic field whose frequency(fext) is close to the fFGL which changes the trajectory of magnetization precession. Since the dc resistance of the STO depends on the trajectory of precession, the effect of injection locking is reflected in the dc resistance of the STO. We can use this phenomenon to detect the fFGL without measuring conventional RF output signal from STO. To demonstrate this experimentally, we used the design of all-in-plane STO which was shown previously as suitable to the MAMR application[3]. We fabricated an STO pillar of around 40nm in diameter and also a co-planar waveguide on top of it to act as an RF antenna that generates an oscillating magnetic field, electrically separated by SiO2. As shown in fig. 1, a bias current and magnetic field were applied to the STO as usual, and the STO underwent oscillation at the suitable values of bias current and magnetic field. Then, we applied a pulsed RF electrical excitation to the co-planar waveguide from a signal generator, with frequency fext and pulsed at 1MHz, and monitored the change of dc resistance between off and on parts of the pulse (ΔR) by using a lock-in amplifier with the pulse frequency as reference. Note that although we measured the resistance oscillation of the STO at a frequency of 1 MHz, the measurement result reflected the dc resistance. Because 1 MHz is much smaller than the STO oscillation frequency of few tens of GHz, the RF resistance change due to the STO oscillation is averaged over several precession cycles of the magnetization. Fig. 2 shows an example of such measurement. Fig. 2(a) shows a typical spectrum of the RF output signal with a tiny peak at 32GHz from STO under dc bias voltage (Vdc) = 50mV and applied magnetic field (Hext) = 1200mT, from which we could not identify whether or not the frequency of this peak corresponds to the actual magnetization precession. Fig. 2(b) shows the value of ΔR as a function of fext for the same values of Vdc and Hext. We clearly detected the change in ΔR as a function of fext, corresponding to injection locking. This informs us of the actual frequency of magnetization precession without measuring the RF output signal. In conclusion, we demonstrated that the dc resistance of the STO changes measurably upon injection locking to an externally applied magnetic field. This can be used to characterize the oscillation frequency of the magnetic field generated by the STO which is already a part of the write head assembly. This will be beneficial for developing an STO-based microwave magnetic field generator for MAMR.

**References:**

[1] J. G. Zhu, X. Zhu, and Y. Tang, IEEE Trans. Magn., vol. 44, no. 1, pp. 125–131, (2008) [2] S. Urazhdin, et al., Phys. Rev. Lett., vol. 105, no. 10, p. 104101, (2010) [3] H. Sepehri-Amin et al., J. Magn. Magn. Mater., vol. 476, pp. 361–370, (2019)

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/dd3c256437e821279ec173fbd1dae0ce.jpg)
Fig. 1 A schematic of our measurement. The STO is connected to a bias tee setup which allows a mix of dc voltage input (Vdc), a high-frequency component and a low-frequency component. The high-frequency component is not to be used in our newly proposed method

![](https://assets.underline.io/uploads/markdown_image/1/image/301904ba6e5647975c3894317f292714.jpg)
Fig. 2 (a) PSD of the RF electrical signal from the STO (b) Change of resistance △R as a function of externally applied RF field frequency fext under the same bias conditions

 Characterizing the oscillation frequency of a spin-torque oscillator (STO) by measuring the change of dc resistance upon injection locking to an external RF magnetic field

Recently, the problem of information storage becomes more serious due to the information explosion. It is widely known that future advances in the areal density of hard disk drives (HDDs) are essentially limited by the magnetic recording trilemma arising from the conflicting requirements between signal-to-noise ratio, writeability, and stability[1]. Energy-assisted magnetic recording (EAMR) such as heat-assisted magnetic recording[2] and microwave-assisted magnetic recording[3] is thought to solve the trilemma by improving writability and enhancing areal recording density. Two-dimensional magnetic recording with shingled writing[4] has also been proposed for increasing track density, which is enabled by partial overlapping of recorded tracks. EAMR combined with SMR is expected to be achieved for the next generation recording schemes of HDDs. On the other hand, servo signal writing is one of serious problems for high areal recording density in hard disks. The self-servo track writing technology, which regenerated from previous written servo patterns, has been adopted during head-disk assembly. However, since the servo signal writing in the present hard disks takes more than a few days per one drives on manufacturing process, and the supply of HDDs can not keep up with the social demand. Magnetic printing with perpendicularly magnetized patterns is a strong candidate for servo track writing with extremely high speed and low cost[5]. However, the recording field strength by utilizing the conventional magnetic printing master medium is not enough to write servo-signals on HDDs with large coercivity for EAMR, and it becomes much weaker as the pattern size or the track pitch decreases. The decrease in the recording field or the field gradient deteriorates the signal-to-noise ratio of servo-signals, and this problem becomes more serious in HDDs which has narrower track width. In order to accurately write servo-signals on HDDs for EAMR combined with SMR, it is essential to enhance the recording field. I propose a new master structure, herein called double magnet master (DMM) medium[6], for magnetic printing technique to improve recording characteristics. Figure 1 shows the schematic illustration of double magnet master medium and recording layer. The DMM medium consists of two magnetic materials with different coercivity. Both materials have perpendicular magnetic anisotropy but exchange interaction between both materials does not act. One material has a coercivity higher than the coercivity of the recording layer, while the other has a lower coercivity. During application of printing field Hp, the magnetization of low coercivity parts (magnet 1) orients to the direction of printing field while high coercivity parts (magnet 2) is not subject to magnetization reversal. As a result, the recording field can be enhanced. The recording field distributions of L/S patterns with pattern length of 10 nm, which corresponds to bit length, were calculated by micromagnetic simulation for conventional perpendicular master (CPM)[5] and DMM media. Tha magnetic spacing is 2 nm in this case. The printing field Hp was varied from 5 kOe to 50 kOe. In CPM, the recording field Hr consists of the magnetostatic field generated from only the magnet 1, while in DMM, the recording field Hr consists of the magnetostatic fields generated from both magnets 1 and 2. The recording field difference ΔHr of about 11 kOe for DMM is twice as that for CPM in various printing field. It was clearly found that the recording field can be enhanced by DMM. Since the magnetostatic field generated from double magnets is effective when the magnetization in the harder magnet does not reverse, the printing field is less than the coercivity of the harder magnet, which is about 73kOe by using L10-FePt patterms as magnet 2. In order to evaluate printing characteristics of CPM and DMM, the printed magnetization distributions of ECC-type EAMR media [7] with the coercivities of 15kOe and 22kOe were calculated. Figure 2 shows printing performance [8] as a function of printing field Hp. The printing performance of 1 means the perfect printing. The optimum printing fields are slightly less than those of the coercivity of recording layer. The printing performances of DMM were dramatically improved comparing with CPM printing characteristics. It was clearly confirmed that the printing performances of DMM are superior to that of CPM. Therefore, the magnetic printing by utilizing DMM is a promising way to realized high-speed servo-track writing for EAMR combined with SMR.

**References:**

[1] H. J. Richter, A. Y. Dobin, J. Magn. Magn. Mater. 287 (2005) 4150. [2] M. H. Kryder, E. C. Gage, T. W. McDaniel, W. A. Challener, R. E. Rottmayer, G. Ju, Y. T. Hsia, M. F. Erden, Proc. IEEE 96 (2008) 1810-1835. [3] J.-G. Zhu, X. Zhu, Y. Tang, IEEE Trans. Magn. 44 Iss. 1 (2008) 125-131. [4] A. Amer, J. Holliday, D. D. E. Long, E. L. Miller, J.-F. Paris, T. Schwarz, IEEE Trans. Magn. 47, Iss. 10 (2011) 3691-3697. [5] N. Sheeda, M. Nakazawa, H. Konishi, T. Komine, and R. Sugita, IEEE Trans. Magn. 45 (2009) 3676-3678. [6] T. Komine, submitted to J. Magn. Magn. Mater. 2021. [7] T. Tanaka, S. Kashiwagi, Y. Kanai, K. Matsuyama, J. Magn. Magn. Mater. 416 (2016) 188-193. [8] T. Komine, T. Murata, Y. Sakaguchi, R. Sugita, IEEE Trans. Magn. 44 Iss. 11 (2008) 3416-3418.

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/d8f8ad00cc0a9b7cd3f35f27da53b2ef.jpg)
Fig.1 Schematic illustration of magnetic printing process by utilizing double magnet master medium.

![](https://assets.underline.io/uploads/markdown_image/1/image/dda58c7ac541a3940c784a3e79047b17.jpg)
Fig.2 Printed magnetization distribution of recording layer and printing performance as a function of printing field.

 Double magnet master media for magnetic printing onto energy-assisted magnetic recording media

The ability to engineer magnetic phase distributions over nanometer length scales is invaluable in the development of spintronic and data storage technologies with enhanced efficiency. Equiatomic FeRh with B2 crystallography is a useful material for local manipulation of magnetic phases, owing to its unusual first-order metamagnetic phase transition (MPT) from antiferromagnetic (AF) to ferromagnetic (FM) ordering, which occurs at ~100 °C [1]. It has been shown that in addition to the thermally driven MPT, there is a disorder-induced MPT from AF to FM ordering, and that with further disordering a second MPT to PM behaviour can be induced [2]. Bennett et al. [3] have demonstrated that by irradiating FeRh with noble gas ions, the disorder-induced MPTs can be exploited to produce a rich variation in the magnetism as a function of depth. In the present work [4], we combine several non-destructive experimental techniques with extensive simulations to identify a direct correspondence between the nature and concentration of induced defects and the resulting modification of the magnetic depth profile in 40 nm FeRh thin films as the fluence of 25 keV Ne+ ions increases. Polarised neutron reflectometry (PNR) was used to determine the magnetisation depth profiles for a range of sample temperatures under an applied field of 1 kOe. The data were fitted to dynamical simulations of the nuclear and magnetic scattering length density (n/mSLD) profiles (Figure 1), which are proportional to the density and in-plane magnetisation, Ms, respectively. The data demonstrate that it is possible to produce a uniform and thermally stable magnetisation with a single-step irradiation, as in Figure 1(b), or to produce a simple bilayer phase structures, as in Figure 1(c) and (d). To elucidate the interplay between the magnetisation depth distributions, the coexistence of magnetic phases, and the ion-induced damage profiles, conversion electron Mössbauer spectroscopy (CEMS) and Doppler broadening variable energy positron annihilation spectroscopy (DB-VEPAS) were performed on each sample. DB-VEPAS allows relative changes of defect concentration to be determined via the line broadening of gamma photons produced from the annihilation of positrons with varying energy Ep. By fitting the DB-VEPAS data, the mean positron diffusion length as a function of increasing fluence is obtained, which is inversely proportional to the concentration of open volume defects. Thus, we demonstrate that the defect concentration monotonically increases with the irradiating fluence. To complement the DB-VEPAS measurements, variable energy positron annihilation lifetime spectroscopy (VEPALS) measurements were also performed on each sample. Here, variations in positron lifetime with increasing fluence signify changes to the size of open volumes; thus, changes in the concentration and the type of defects are determined. By comparing the measured data to theoretical lifetime values calculated using density functional theory (DFT), we identify the predominant defect type for each irradiation condition with depth sensitivity. The PNR, CEMS, and positron data allow the effect of increasing fluence to be explained as shown schematically in Figure 2, wherein the onset of FM and PM phases are associated with the emergence of static disorder and monovacancy defects respectively. In the as-grown case, there is a substantial concentration of monovacancy defects. At the lowest value of the fluence (7.5 × 1013 ions cm-2), the defect concentration increases and sub-unit cell atomic displacements begin to dominate. The generation of static disorder continues as the fluence is increased to 3 × 1014 ions cm-2, with no change to the predominant defect type. On increasing the fluence to 6 × 1014 ions cm-2, atomic displacements give way to monovacancies, with some static disorder remaining at the substrate interface. For all samples, our data are well described assuming AF, FM, and PM phases only, and do not present evidence for states such as spin glasses or magnetic dead layers. This feature, coupled to the fact that the mSLD profiles of the irradiated samples comprise relatively few sublayers with discrete magnetisations which are stable with respect to increasing temperature, highlights the effectiveness of sensitive and controlled disordering of ordered alloy precursors for producing technologically useful magnetic configurations. Our results provide the basis for the patterning of single layer or multilayer structures that incorporate FeRh to create devices on sub-100 nm length scales with AF and FM order in close physical proximity.

**References:**

[1] M. Fallot, Ann. Phys., vol. 10, pp. 291–332 (1938) [2] A. Iwase et al., Nucl. Instrum. Methods Phys. Res., vol. 256, pp. 429–433 (2007) [3] S. P. Bennett et al., Mater. Res. Lett., vol. 6, pp. 106–112 (2018) [4] W. Griggs et al., APL Mater., vol. 8, p. 121103 (2020)

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/69319fe9359c4803c0c33b75650a7a95.jpg)
Fig. 1. (a)-(d) Magnetic (coloured) and nuclear (grey) scattering length density (mSLD and nSLD) profiles corresponding to the fitted PNR data for each sample across the range of measured temperatures.

![](https://assets.underline.io/uploads/markdown_image/1/image/85121693078388560eeaf904d351ce49.png)
Fig. 2. The effect of increasing the irradiating fluence on the depth-dependent magnetisation, magnetic phase fractions, and concentration and type of structural defects is illustrated.

 Depth Selective Magnetic Phase Coexistence in FeRh Thin Films

As the current recording media for the heat assisted magnetic recording (HAMR), the granular L10-FePt thin film needs to have uniform grain size with small pitch distance and well-defined grain boundaries. To achieve the targeted microstructure with proper L10 ordering using sputtering techniques, FePt along with grain boundary materials (segregants) are co-deposited on MgO underlayers at elevated temperature. The columnar growth of FePt is desired to have sufficient grain volume without grain size expansion. The fabrication of grains at small grain pitch with well-defined grain boundaries has been proven to be challenging.[1] Over the past decade, segregant materials, such as carbon, have attracted extensive attention for producing isolated FePt grains with well-defined grain boundaries but has experienced difficulty at reducing grain pitch distance and making grains with large aspect ratio. Though the columnar growth of FePt can be achieved by using some oxide segregant materials, such as TaOx[2] and TiOx[3], the resulting films always show poor L10 ordering and poor magnetic properties. Proper ordering of FePt can be achieved with SiOx[4], but the resulting microstructure tends to be maze-like. Therefore, exploiting multilayer deposition is a promising strategy to achieve the desired microstructure with proper ordering and magnetic properties. In this study, we utilize a dual-layer strategy of FePt-BN/FePt-SiOx on polycrystalline MgO underlayer in order to obtain the desired media microstructure with good L10 ordering and magnetic properties. The film stack of Ta(5)|Cr(40)|MgO(6)|Media(6.5) was deposited on Lotus NXT glass substrate with the numbers in parentheses in nm. Lotus NXT glass is one of the promising glass substrates for high-temperature deposition of HAMR media. The Ta and Cr layers were DC sputtered at room temperature and 280°C, respectively. The samples were subsequently annealed at 650°C for 1 hour in order to promote the texture and grain size of Cr buffer layer. The MgO underlayer was RF deposited on Cr buffer layer after the sample cooled down to room temperature. There are two magnetic layers in the media layer, where FePt was co-sputtered with BN and SiOx, respectively, for the first (M1) and the second (M2) magnetic layer. The deposition temperatures of the M1 layer and the M2 layer were fixed at 700°C and 550°C, respectively. From the microstructural point of view, the main objective of the M1 layer is to obtain well-isolated FePt grains to set the grain center-to-center pitch distance, whereas the M2 layer aims to suppress the lateral growth and to promote the columnar growth of the FePt grains. Fig. 1 (a) shows the in-plane HAADF image of the film stack Lotus NXT glass |Ta(5)|Cr(40)|MgO(6)|Media(6.5) with its (b) grain size distribution profile and (c) grain center-to-center pitch distance distribution profile. The figure demonstrates well-isolated FePt grains with mean grain size of 6.3 nm and mean pitch distance of 8.3 nm. As shown in the XRD pattern of this media in Fig. 1 (d), there are only (001) and (002) peaks from FePt, indicating the c-axis of FePt grains are predominantly perpendicular to the film plane. The broad peak at around 23.5° corresponds to the amorphous peak from Lotus NXT glass substrate. The order parameter (S) considering the relative integrated intensity ratio between (001) and (002), I001/I002, with the absorption factor and Lorentz factor is 0.7, suggesting good chemical ordering in FePt grains. Fig. 1 (e) shows the perpendicular and in-plane hysteresis loops of this media. The perpendicular coercivity (Hc⊥) is 26 kOe. The in-plane loop shows a small opening with the coercivity (Hc//) of 2.9 kOe and the normalized remanence magnetization (Mr//) of 0.08. The remanent ratio, which is Mr// /Mr⊥, is 0.09, suggesting a minor in-plane component of FePt grains in the sample. By exploiting the dual-layer structure of FePt-BN/FePt-SiOx, we have demonstrated well-isolated columnar FePt grains with small mean pitch distance of 8.3 nm. The XRD pattern shows that the grains are well-textured and highly ordered. As a result, the perpendicular coercivity reaches 26 kOe. The utilization of multilayer deposition strategy with segregants BN and SiOx provides a promising technique to achieve desired HAMR media microstructure without sacrificing FePt chemical ordering and magnetic properties.

**References:**

[1] Y. Kubota et al., IEEE Trans. Magn., vol. 54, no. 11, p. 3201206, 2018 [2] B. Lim et al., J. Appl. Phys., vol. 105, no. 7, p. 07A730, 2009 [3] B. S. D. C. S. Varaprasad et al., IEEE Trans. Magn., vol. 49, no. 2, p. 718–722, 2013 [4] H. Ho et al., Appl. Phys. Lett., vol. 102, no. 11, p. 112411, 2013

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/44d762af9dd6d7af7c27e30c471fccc1.jpg)
Fig. 1 (a) In-plane HAADF image of the film stack Lotus NXT glass |Ta(5)|Cr(40)|MgO(6)|FePt-38vol.% BN(2)|FePt-35vol.%SiOx with its corresponding (b) grain size distribution and (c) grain center-to-center pitch distance. Its XRD pattern and magnetic properties are in (d) and (e) respectively.

 Fabrication of FePt-BN/FePt-SiOx dual-layer structure for HAMR media on Corning Lotus™ NXT glass substrate

I. INTRODUCTION In recent years, spin-torque transfer magnetic random access memory (STT-MRAM) has received tremendous attention due to its low switching energy, fast read/write speed, and high scalability [1]. However, thermal fluctuation and process variation have a detrimental effect on the data recovery of STT-MRAM, leading to both read errors and write errors [2]. Error correction codes (ECCs), such as the Hamming codes and Bose-Chaudhuri-Hoquenghem (BCH) codes have been employed to improve the reliability of STT-MRAM [2, 3]. Meanwhile, the channel quantizer that quantizes the signal read back from the STT-MRAM cell is critical to support the above ECCs, since high-precision analog-to-digital converters (ADCs) are not applicable for STT-MRAM. Various criteria have been proposed for designing the quantizer for STT-MRAM, including the Maximizing-Mutual-Information (MMI) criterion, the Maximizing-Cutoff-Rate (MCR) criterion, and the Optimizing- Polyanskiy-Poor-Verdu -Bound (PPVB) criterion [4]. However, the final choice of the quantization scheme is still relying on the decoding error rate performance of ECCs. It is typically evaluated by using computer simulations, which are too slow to reach the target error rate level of 10-12 or below of data storage systems. In this work, we propose a union bound analysis which can accurately predict the word error rates (WERs) of ECCs with maximum-likelihood (ML) decoding over the quantized STT-MRAM channel. The derived bound provides a theoretical tool for comparing different quantization schemes at the targeted low error rate levels of data storage systems. Moreover, we also present a new criterion to design the quantizer by minimizing the WERs of ECC decoding that are obtained from the union bound analysis. Numerical results show that the union bound optimized quantizer can achieve better WER performance than the state-of-art quantizers for STT-MRAM. II. UNION BOUND ANALYSIS AND CHANNEL QUANTIZATION The union bound analysis is a theoretical technique that is used for bounding the error rate of ML decoding of ECCs. It was mainly developed for codes over symmetric channels [5]. In [6] and [7], more sophisticated union bound techniques were proposed for simple asymmetric channels, i.e. the Z-channel and the general binary asymmetric channel (BAC). In this work, we consider the quantized STT-MRAM channel with both write errors and read errors, which can be modelled as a concatenation of a BAC and a binary-input multiple-output discrete memoryless channel. Our proposed union bound analysis first utilizes a function of a predefined multi-dimensional distance between two codewords to calculate the pairwise error probability (PEP) [7] of the quantized STT-MRAM channel. Note that this is different from the cases of the symmetric channels where the PEP is uniquely determined by the Hamming distance of the code. Based on the PEP, we can derive a union bound of the WER of ML decoding as a function of the weight distributions of ECC. Next, we further propose to use the WER obtained from the above union bound analysis as the criterion to design the channel quantizer of STT-MRAM. We also develop effective differential evolution algorithm to determine the optimum quantization boundaries for the multi-bit quantizers. Since our proposed union bound optimized quantizer takes into consideration the weight spectrum of ECC, it can achieve better error rate performance than the prior-art quantizers. III. NUMERICAL AND SIMULATION RESULTS The system parameters of STT-MRAM are taken from [2]. In the quantized STT-MRAM channel model, we assume the write error rate of 10-5, and vary the mean normalized resistance spreads σ0/μ0=σ1/μ1 of the low and high resistances to account for the influence of different process variations. The (72,64) extended Hamming code [3] is adopted to validate our analysis, and the simulated ML decoding performance is obtained using the Bahl, Cocke, Jelinek and Raviv (BCJR) decoder [5]. In Fig.1, we compare the decoding WERs provided by the union bound with those achieved by the BCJR decoder through simulations, with the proposed union bound optimized quantizer with different number of quantization bits q. Observe that there is a noticeable gap between the union bound and simulated WERs for the case of q=1. However, with two or more quantization bits, the performance gap becomes negligible, and when the WER is less than 10-6, the union bound is almost identical to the simulated WER. In Fig. 2, we fix the quantization bit to be q=2, and compare the WER performance with different types of quantizers. A nice agreement is observed again between the WERs predicted by the union bound and those obtained from simulations, for the various types of quantizers. Moreover, our proposed union bound optimized quantizer outperforms all the other quantizers. There is a larger performance improvement over the MMI quantizer, which is most widely adopted in the literature. This demonstrates the potential of our proposed union bound analysis in predicting the decoding error rate performance and in guiding the design of channel quantizer for STT-MRAM. Acknowledgement: This work is supported by RIE2020 Advanced Manufacturing and Engineering (AME) programmatic grant A18A6b0057 and Singapore MOE Tier 2 fund MOE2019-T2-2-123.

**References:**

[1] M. Marinella, “The future of memory,” in Proc. Aerospace Conf, Big Sky, MN, USA, Mar. 2012, pp. 1–11. [2] K. Cai, and K. A. S. Immink, “Cascaded channel modeling, analysis, and hybrid decoding for spin-torque transfer magnetic random access memory (STT-MRAM),” IEEE Trans. Magn., vol. 53, no. 11, pp. 1–11, 2017. [3] K. Cai, Z. Qin, and B. Chen, “Channel capacity and soft-decision decoding of LDPC codes for spin-torque transfer magnetic random access memory (STT-MRAM),” in Proc. Int. Conf. Comput., Netw., Commun. (ICNC), San Diego, CA, USA, Jan. 2013, pp. 550–554. [4] Z. Mei, K. Cai and L. Shi, “Information Theoretic Bounds Based Channel Quantization Design for Emerging Memories,” in Proc. IEEE Information Theory Workshop (ITW), Guang Zhou, China, Nov. 2018, pp. 1–5. [5]W. E. Ryan and S. Lin, Channel Codes: Classical and Modern, Cambridge, Cambridge University Press, 2009. [6] A. Barbero, P. Ellingsen, S. Spinsante, and O. Ytrehus, “Maximum likelihood decoding of codes on the Z-channel,” in Proc. IEEE International Conference on Communications (ICC), Istanbul, Turkey, Jun. 2006, pp. 1–5. [7] G. Song, K. Cai and J. Cheng, “A union bound analysis for codes over binary asymmetric channels,” in Proc. IEEE International Conference on Communications (ICC), Paris, France, May. 2017, pp. 1–5.

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/7687adc035569fbf9f2d0102c1f4556f.jpg)
Fig. 1. Comparison between analytical and simulation WERs, with union bound optimized quantizer with q= 1,2,3.

![](https://assets.underline.io/uploads/markdown_image/1/image/ff97a646ffd6d77b07f4e57f875be682.jpg)
Fig. 2. Comparison between analytical and simulation WERs with different types of quantizers, for q=2.

 Union Bound Analysis for Spin-Torque Transfer Magnetic Random Access Memory with Channel Quantization

I. Introduction Track misregistration (TMR) and two-dimensional (2-D) interference that consists of an inter-track interference (ITI) and inter-symbol interference (ISI) are the significant problems in an ultra-high bit-patterned media recording system (BPMR) which easily degrade the bit-error-rate (BER) performance of the system, because of the extremely narrow distance between bit-islands. It very challenges to control the reader to fly above the center of the desired track. Normally the TMR effect can be controlled by servomechanism [1] which required extra area to store the redundancy bits for tracking the data track. Instead, several works use only readback signals to estimate and correct the TMR effect [2-4]. For example, we proposed a TMR prediction method by using the energy ratio of readback signals, and then the TMR effect can be dealt with utilizing the appropriate 2-D equalizers together with the use of soft-information adjustment (SIA) technique [5]. II. Proposed Methods To improve the TMR estimation accuracy and BER performances of BPMR system; therefore, we propose to use the neural network models called a multi-layer perceptron (MLP) in both TMR level prediction and TMR effect correction processes. We consider the dual-reader two-track reading under BPMR media at an areal density (AD) of 3.0 Tbits/in2, which track-pitch, Tz and bit-period, Tx are set to be 14.5 nm. In the reading process, 2 readers simultaneously read the medium of upper- and lower-tracks, while the sidetracks located above and below two main tracks will be also read, which denoted as r-1,k, r0,k, r1,k, and r2,k. Then, 4 readback signals are equalized by 1-D (one-dimensional) equalizers which are designed according to signal with no TMR effect. The equalized data sequences will then be multiplied with their absolute value for each data sample to extend their amplitude before sending them to the TMR level prediction and TMR effect correction processes, respectively. In TMR level prediction, four modified readback signals, u-1,k, u0,k, u1,k, and u2,k with the size of 4×32,760 bits will then be grouped into 4×3 bits in each group before sending to the TMR level prediction process. That means each considered group consists of 12 data bits and so there are 10,920 groups in one set of readback sequence, the number group represented by j ∈ {1, 2, 3, …, 10920}. Fig. 1(a) shows the structure of the MLP TMR estimator which consists of 1 input layer and 12 input nodes, 5 hidden layers consist of 100 hidden nodes, and 1 output layer with a single node. This MLP TMR estimator receives a group of readback signals at a time for predicting the estimated TMR level, TMR'j of the jth group. Then, the estimated TMR levels of all groups were averaged to obtain the estimated TMR level, TMR' of the entire sequence. Here, the accuracy percentage of TMR' is measured as accuracy (%) = 100 – (|TMR'-TMR| / TMR × 100) [5]. we also propose to use the MLP data detector which may be employed instead of the conventional detectors-based PRML system. Fig. 1(b) shows the structure of the MLP data detector, which is identical to MLP TMR estimator except for the number of input and output nodes. The MLP data detector receives a group of modified readbacks similar to the MLP TMR estimator, but it also receives an additional estimated TMR level, TMR' that was produced from MLP TMR estimator. Here, it is fixed as one of the input data of the MLP data detector. Finally, we can obtain the estimated user bits of both upper- and lower-tracks, i.e., a'0,k and a'1,k, in the size of 2×3 bits for all 10,920 groups. In this study, the leaky rectified linear unit (ReLU) is used as an activation function for the hidden nodes and linear function in the output nodes for both MLP models. The linear regressions are then adopted for calculating their outputs. The Adam algorithm is used to optimize the weights and bias of both MLP models. III. Results and Discussion In this study, we start with the consideration of an accuracy percentage of TMR level estimation in various TMR and SNR levels where the position jitter is given to be 0%. The TMR level is varied from 0% to 15% of a track pitch as mentioned above. The signal-to-noise ratio (SNR) that is defined as 10log10(1/σ2), where σ is a standard deviation of electronics noise, is first considered at 10, 15, 20, and 25 dBs. We found that the proposed MLP TMR estimator can provide an accuracy percentage higher than 80% for all 4 considered SNRs and all considered TMR levels. Especially, at the higher SNR, e.g., 15, 20, and 25 dBs, we can get more than 90% accuracy for all TMR levels. Moreover, we also investigate the BER of our proposed MLP detector, which is compared with the convention partial response maximum likelihood (PRML) system that performs based on 2-D equalizers and 2-D Viterbi detectors together with the SIA technique [5] where it is represented with “conventional”. As shown in Fig. 2, it seems that at higher SNR, our proposed system can provide better performances for all TMR levels. For example, at BER = 10-4 and TMR level is 0%, our proposed system can provide more than 5 dB over the conventional system. The system performance implies that it is possible to use the MLP neural networks for signal processing in BPMR system.

**References:**

[1] Y. B. Chang, D. K. Park, and N. C. Park, IEEE Trans. Magn., Vol. 38, no. 2, p. 1441 (2002) [2] C. Warisarn, IEEE Trans. on Magn., Vol. 55, no. 7, p. 1 (2019) [3] W. Busyatras, C. Warisarn, and Y. Okamoto, AIP Adv., Vol. 7, p. 056501 (2016) [4] S. Han, G. Kong, and S. Choi, IEEE Transactions on Magnetics, Vol. 55, no. 7, p. 1 (2019) [5] K. Kankhunthod, W. Busyatras, and C. Warisarn, 17th ECTI-CON., p. 735 (2020)

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/09de0a5a9c8f4396146c47ebd77277d5.jpg)
Fig. 1: MLP structures of the TMR level estimator and TMR detector.

![](https://assets.underline.io/uploads/markdown_image/1/image/fd1a13ccd5cbab14c903baba723d02af.jpg)
Fig. 2: BER performance comparison between the proposed MLP based TMR correction method versus the conventional method at the AD of 3.0 Tbits/in2, here the position jitter is fixed to be 0%.

Enable TDMR Gain with Convolution Neural Network Based Machine Learning Algorithm

Downloads

Next from INTERMAG 2021

Thermal gradient driven dynamics of Neel skyrmions in a nanoracetrack

Similar lecture

Coexistence of distinct skyrmion phases observed in hybrid ferromagnetic/ferrimagnetic multilayers

Enable TDMR Gain with Convolution Neural Network Based Machine Learning Algorithm

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from INTERMAG 2021

Thermal gradient driven dynamics of Neel skyrmions in a nanoracetrack

Similar lecture

Coexistence of distinct skyrmion phases observed in hybrid ferromagnetic/ferrimagnetic multilayers

Downloads