United States

We report a novel delayed-electroforming process in Metal-Insulator-Metal Au/Ti/SiOx/Mo ReRAM devices; this new process allows more energy-efficient switching. For this novel process to occur, a highly resistive pristine device must be stressed under a fixed high voltage for a varying amount of time. The stress time of a device is determined by the amount of charge going through the oxide during stressing. Once a fixed amount of charge has gone through the device, a small detection voltage of 10mV is applied. The device continues to show high resistance (~MΩ) after stressing, and abruptly reaches a low resistance state (~100Ω) after a delay. We investigated the effect of charge injection during stressing, and found good correlation between charge injected during stressing and the delay time after stressing. Additionally, we found that we can modulate conductance of the formed devices by varying amount of charge injection. Based on these results, we propose a charge assisted electroforming model to explain this novel phenomenon.

Materials Challenges for Memory

A Charge-Assisted Delayed-Electroforming Phenomenon in Silicon Oxide

silicon oxide

electroforming

reram


![](https://assets.underline.io/uploads/markdown_image/1/image/2e10c515564e9ec3287ee4c664011c20.png)

The energy cost of storing and moving data is enormous, and is growing at such a rapid pace that within a few years the power demand may overwhelm some developed nations. Non-volatile memory (NVM) and neuromorphic computing are enabling critical advances in this area, but there are many scientific and technological challenges. This conference will cover the materials challenges of different memory technologies, from the fundamental processes and mechanisms, to industry perspectives on materials specifications and manufacturing.

poster

Logic-in-memory architectures based on resistive random access memory (RRAM) technology and the material implication (IMPLY) logic are a promising solution for the development of energy efficient computing hardware for edge computing applications. However, RRAM devices nonidealities, such as cycle-to-cycle variability and random telegraph noise, and circuit-related issues such as logic state degradation, strongly affect the reliability of these circuits and complicate their design. In this, work we use a physics-based RRAM compact model which reproduces the effect of device nonidealities (i.e., variability, Random Telegraph Noise, self-heating) and that was calibrated on three RRAM technologies from the literature to study the reliability and performance of a conventional and a recently proposed smart IMPLY architectures. Finally, we underline the importance of the physics-based RRAM compact model as a tool for implementing device-circuit co-design strategies.

Device-circuit co-design of logic-in-memory circuits using a physics-based RRAM compact model

Electrospinning can be used to produce nanofibers from magnetic materials, forming a magnetic network which may be applied in neuromorphic computing. Besides the influence of the bending radii on the magnetization reversal in such nanofibers [1], the often occurring beads along the fibers have not yet been investigated in detail. Such beads can be assumed to support domain wall formation due to a reduction of the relative impact of the shape anisotropy, in this way influencing magnetization reversal along the fiber. 
Here, we show results of simulations of magnetization reversal processes in iron, nickel, cobalt and magnetite fibers with a bead in the middle, produced in three typical dimensions gained by electrospinning [2]. Mostly, magnetization reversal occurs via a vortex state, for all materials and dimensions. Only some angular orientations of the external magnetic field result in double-vortex or meander states, usually followed by a vortex state. The simulations emphasize the impact of beads along electrospun magnetic nanofibers, making magnetization reversal highly reliable and thus such structures useful for data storage and transport in neuromorphic computing and other applications.

References
[1] T. Blachowicz, A. Ehrmann, Magnetization reversal in bent nanofibers of different cross-sections, J. Appl. Phys. 124, 152112 (2018).
[2] F. Amini, T. Blachowicz, A. Ehrmann, Systematic study of magnetization reversal in beaded fibers from different magnetic materials, J. Magn. Magn. Mater. 529, 167855 (2021)

Magnetization reversal in nanofibers with beads from different magnetic materials

Real-time tracking of high-speed objects in cognitive tasks is challenging in the present artificial intelligence techniques because the data processing and computation are time consuming, resulting in impeditive time delays. A brain-inspired continuous attractor neural network (CANN) can be used to track fast moving targets, where the time delays are intrinsically compensated if the dynamical synapses in the network have short-term plasticity. Here, we show that synapses with short-term depression can be realized by a magnetic tunnel junction, which perfectly reproduces the dynamics of the synaptic weight in a widely applied mathematical model. Then, these dynamical synapses are incorporated into one-dimensional and two-dimensional CANNs, which are demonstrated to have the ability to predict a moving object via micromagnetic simulations. This portable spintronics-based hardware for neuromorphic computing needs no training and is therefore very promising for the tracking technology of moving targets. [1]
[1] Q. Zheng, X. Zhu, Y. Mi, Z. Yuan, and K. Xia, Physical Review Applied 14, 044060 (2020).

Anticipative Tracking with the Short-Term Synaptic Plasticity of Spintronic Devices

A resistive switching ZnO homojunction diode fabricated with a stack of two dissimilar ZnO thin film shows promising characteristics necessary to address the crosstalk [1] problem in a cross-bar memory array. The device has a characteristic IV-hysteresis loop in forward bias, in conjunction with its pn junction characteristics. In the ac frequency range of 10 Hz to 15 kHz, the area inside the hysteresis loop initially increased, reaching a maximum at 1 kHz, and then decreases with increasing frequency. The maximum in the area of the hysteresis loop corresponds to the accumulation of space charge at the interface. The resistivity of the device was found to vary from 5 kΩ to 100 kΩ in the bias range 0V - 10V and in the range of 106 Ω when reverse biased the ratio of High Resistive State (HRS) to Low Resistive State (LRS) [2] has its maximum (~2.4) at 1 kHz ac signal at 3V.


Addressing crosstalk in cross-bar memory arrays with a resistive switching ZnO homojunction diode

Observing the domain structure of antiferromagnetic materials is notoriously difficult, and in the non-collinear antiferromagnet Mn3NiN eight arrangements of the local magnetic moments are possible in the unit cell. Yet for potential applications in antiferromagnetic spintronics it is essential to prepare known domain structures to control the properties of the devices. We grow Mn3NiN films 50nm thick on STO, pattern the films into Hall bars and cool below TN = 220K into the antiferromagnetic state. We then use a laser to generate a local in-plane heat gradient and detect the voltage generated due to the anomalous Nernst effect (ANE) - a manifestation of the Berry curvature of the material. From the sign of the voltage signal we can distinguish domains 2μm in diameter and can saturate the device to a single state using an applied magnetic field. We also show a form of heat assisted writing - by increasing the laser power and reversing the field we are able to write a domain with opposite sign of the ANE.

Using the anomalous Nernst effect to image domains in non collinear antiferromagnetic Mn3NiN thin films

Wake-up effect is a major issue for ferroelectric HfO2-based memory devices. Here, two TiN/Hf0.5Zr0.5O2/TiN structures deposited by magnetron sputtering on silicon are compared. The maximum remanant polarization is higher than 21µC/cm² for both samples, a strong difference is observed in the electrical behavior. For the mesa sample, the difference between maximal and initial Pr is only 3µC/cm² whereas it is around 14µC/cm² in the non-mesa case. We discuss the root causes of these behaviors in light of structural results. Various applications have been suggested for fluorite-structure ferroelectrics due to their advantages over the conventional perovskite-structure ferroelectrics. In this presentation we will focus on (Hf,Zr)O2 (HZO) thin films deposition for the capacitor of Ferroelectric Random Access Memories (FRAM) in the 1Transitor-1Capacitor (1T-1C) model. (Hf,Zr)O2 thin films are studied to either fully understand the stabilization of the ferroelectric phase (f-phase) or to fit with industrial requirements. Samples are stacks of Si/TiN/Hf0.5Zr0.5O/TiN/Pt. All materials are grown by sputtering at room temperature following by a rapid thermal annealing at 450°C during 30 seconds under N2 atmosphere. The samples are called NM, and M. NM and M refers to two different architectures, respectively non-mesa and mesa structures. The description and size of NM and M samples is given in figure 1 for each sample. We report the fabrication of two samples deposited by magnetron sputtering with excellent performances, quite similar to samples deposited by Atomic Layer Deposition. Pr values are among the highest for samples deposited by sputtering. Although the N-sample and NM-samples show very close Pr values, the two samples show completely different electrical behaviors. During cycling, the increase of Pr value for the NM-sample is more than an order of magnitude higher than the M-sample. It is accompanied by a decrease of the endurance which is two order of magnitude higher for the NM-sample than for the M-sample. As electrical behaviors are not the same, for low stress conditions M-sample has a higher Pr value during cycling whereas for high stress conditions NM-sample has a higher Pr value during cycling. As a matter of fact, it has been proven that maximum Pr values are more sensitive to stress conditions than the structures themselves. The origins of the different electrical behaviors come from the micro-crystalline structures of the two samples, according to GIXRD results. The crystallization takes place during the annealing step. During annealing, M-sample is built with a TiN TE fully covering the HZO layer whereas the TiN covers only partially the HZO layer in case of the NM-sample. It induces different stress states which lead to two different micro-crystalline patterning. The M-sample shows no monoclinic peak, whereas the NM-sample shows many monoclinic orientations. It can explain the huge reduction of the wake-up effect.

Reduction of HfZrO2 capacitor wake-up effect

Artificial neural networks employing the biomimetic functions of memristors and crossbar array architecture have been actively explored and designed for use in energy‐efficient neuromorphic computing. In particular, a memristor with a crossbar array architecture is more efficiently able to perform vector-matrix multiplication compared to conventional CMOS hardware. In this process, a memristor with certain characteristics is required, such as ohmic behavior in all resistance states and symmetry/linearity in its long-term potentiation/depression behavior. In this study, to improve these characteristics, we proposed and demonstrated a Li-based composite memristor (LCM) that consists of three phases: Li-doped TiO2 as a Li reservoir, Li4Ti5O12 as the insulating phase, and Li7Ti5O12, as the metallic phase. In the LCM, resistance switching property was induced by adjusting the ratio of the metallic to the insulating phase. The phase formation and coexistence of TiO2, Li4Ti5O12, and Li7Ti5O12 were confirmed using X-ray diffraction patterns and Raman spectral analysis. The proposed LCM exhibited symmetric and gradual switching behavior for both set and reset switching in a full I-V sweep because the driving force leading to the increase/decrease in the ratio of the metallic phase inside the thin film during set and reset switching was the same. The proposed LCM was thus demonstrated to be promising for improving both the ohmic behavior of synapse devices and the linearity and symmetry of long-term potentiation/depression behavior required for artificial neural networks. 
Acknowledgments: This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (2019R1F1A1059637) and the Ministry of Trade, Industry & Energy (MOTIE, Korea) under the Industrial Strategic Technology Development Program. No. 10068075, 'Development of Mott-transition-based forming-less non-volatile resistive switching memory & array'.


Li-Based Composite Memristor for Artificial Neural Networks

Electrochemical devices allowing to modulate their resistive state through ionic motion upon electric bias are known as memristors. Currently, memristors are positioned as a major candidate to overcome the bottlenecks in current electronic-based transistors in terms of downscaling capabilities and energy consumption. The vast majority of memristor are based on two types of ions: either oxygen vacancy migration, in the so-called Valence Change Memories (VCM), or a metal cation, usually Ag+ and Cu2+, in the so-called Electrochemical Metallization Cells (ECM). Despite the excellent performance of both systems, a widespread implementation of oxygen-based memristors in today's integrated circuits is delayed due to the need to address cycle-to-cycle and device-to-device variabilities while circumventing electroforming, which are inherent issues associated to the filamentary nature of the switching mechanism. Recently, Li-ionic memories are emerging as an alternative, given the higher diffusivity of Li+ when compared to oxygen, and the ability of Li-oxides to accumulate and deplete lithium at the interfaces and bulk. The scarce literature regarding Li-based memristors focuses mainly on well-known high-voltage cathode battery materials like LixCoO2, rather sluggish conductor as LiLaTiO3or magnetic spinels like LiFe5O8.1,2 Performance-wise, Li-based oxides have shown promising applications in neuromorphic computing thanks to a close-to-analog switching response. Despite their promise, it remains unclear how to design the defect chemistry and control the switching behavior of Li-based materials towards their performances. For instance, the material class of Lithium titanate compounds present intrinsic properties that could be beneficial to implement for a Li-operated memristor: a) the existence of a metal-insulator transition upon lithium intercalation, concomitant with a phase transition and valence change in the Ti cation, b) the spinel to rock-salt phase transition occurs with a neglectable volume change and c) there is a large difference in electronic conductivity and diffusivity between the two phases. In this work,3 we first discuss the iono-electronic structure requirements to define Li-oxides and select their chemistries for resistive switching applications and neuromorphic computing. Second, we exemplify the resistive switching capabilities of these systems for different lithium stoichiometries and discuss current challenges and tricks in processing to make these. Third, we conceptually discuss how the Li-stroichiometry can be used as a knob to alter switching performance to be either linear or non-linear for neural computation, and from a material perspective how models for that can be thought of. To exemplify this, we will report most recent results on non-volatile, and non-filamentary bipolar resistive switching characteristics of lithium titanates compounds, Li4+3xTi5O12, that can be operated either for deep or synaptic neural networks applications. Here, we employed a recently proposed strategy to overcome lithium loss during thin film deposition and finely control the final lithiation degree of the films4 to create a stoichiometrically lithiated Li4Ti5O12 spinel phase and a highly lithiated Li7Ti5O12 rock- salt phase memristive devices. Changing the Li-content from a stoichiometrically lithiated spinel phase to a highly lithiated rock-salt phase results in the capability to tune the performance in a wide range in terms of accessible resistance window (from ratios of 102 to 106 at low voltage operation, respectively), switching symmetry (from highly asymmetric to symmetric behavior, respectively) and retention (from few minutes up to 105 s at room temperature, respectively), among others. In other words, controlling the lithiation degree might offer a suitable path to reduce stochasticity from which current filamentary memristive devices inherently suffer, mainly due to the difficulties to control the amount of vacancies generated, and paves a way to further control of ionic migration for novel iono-electronic devices in neuromorphic computation.

Neuromorphic Computing with Li-based Battery Materials

The generation, detection, and manipulation of pure spin currents have attracted considerable attention in the field of spintronics owing to their rich physical phenomena and potential applications [1-2]. The most convenient method for generating and detecting the pure spin currents is to utilize the spin Hall magnetoresistance (SMR) [3-4]. In this letter, a novel structure consisting of two heavy metals that delivers competing spin currents of opposite spin indices is presented. We studied spin Hall effect in Ta-Pt alloys with various Ta composition for Ta/Pt/Ni80Fe20/ MgO (001) heterostructures by means of SMR. To determine the relative contributions of Pt and Ta, we measured a series of samples of Ta（tTa=0-3nm）/Pt（3nm）/Ni80Fe20/MgO (001) with a constant Pt (3) layer and various thicknesses of the Ta layer. Low resistivity, flat, and highly (111)-textured Pt-Ta alloy films are synthesized. It is found the SMR ratio is enhanced with a Ta capping layer at an optimum composition of x between 0.5nm to 1nm. The spin accumulation enhancement at the Pt/Ni80Fe20 interface after capping with a Ta layer, which functions as a spin sink at small thickness, is responsible for the SMR ratio enhancement, even though Ta and Pt have opposite spin Hall angels. Our results not only demonstrate a viable scheme for multiple spin currents, but also expand a new avenue for switching a perpendicular ferromagnetic layer by competing spin currents.

Enhancement of the spin Hall magnetoresistance in Ta/Pt/Ni80Fe20/MgO(100) heterostructures by competing spin currents

Neuromorphic computing architectures have superior properties over conventional computers in solving a number of problems such as image recognition, combinatorial optimization or prediction. As a concept, neural networks have been proven to be fast and flexible. However, their implementation in conventional digital architecture requires large amount of resources, as subsequent multiplication and addition operations need to be performed for each input of a single neuron. We present a complete architecture for neuromorphic computing including: a magnetic tunel junction (MTJ) based multi-bit cell forming a programmable synaptic weight, an electronic neuron and the artificial neural network itself. We demonstrate a successful operation of such network by presenting simulation results of a representative problem solution.


Premium content

Downloads

Next from Materials Challenges for Memory

Device-circuit co-design of logic-in-memory circuits using a physics-based RRAM compact model

Similar lecture

Interface-controlled Resistive Switching in a-InGaZnO Memristors